Backdoor Attacks under Lossy Compression: From Failure to Reactivation and Adaptation

arXiv:2603.13864 · cs.CR, cs.CV · Submitted 2026-03-14 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Security Radio. Generated commentary on the latest security and cryptography papers.

Nadia: I'm Nadia.

Elias: And I'm Elias, here with Priya.

Nadia: We're looking at the paper: "Backdoor Attacks under Lossy Compression".

Elias: Here's the gist of the paper. Real-world backdoor attacks often require poisoned datasets to be stored and transmitted before they are used to compromise deep learning systems.

Nadia: The paper highlights the necessity of explicitly accounting for lossy compression in backdoor attacks.

Elias: The core problem addressed is the failure of invisible triggers under lossy compression.

Nadia: Invisible triggers embedded in RGB images may not survive the conversion from RGB data to compressed binary bitstreams and back to decompressed RGB images.

Paper summary: Nadia: So wrapping up this discussion on "Backdoor Attacks under Lossy Compression: From Failure to Reactivation and Adaptation," the paper really hammers home that we need to explicitly account for lossy compression when designing data poisoning attacks in deep learning systems.

Elias: It moves the focus from just creating a trigger in an image to ensuring that trigger survives the whole storage and transmission pipeline, which is a crucial practical consideration >

Priya: The authors propose Universal Attack Reactivation and Compression-Adapted Attack as specific ways to deal with that, showing how we can reactivate degraded triggers or construct new ones adapted to the compression process >

Nadia: CAA is presented as a method that performs very well, maintaining high attack success rates and stealthiness even under JPEG re-compression, which suggests it's quite generalizable across different codecs >

Elias: It’s important to remember the limitation they acknowledge: they focus on revealing and constructing these compression aware backdoor attacks but don't propose a dedicated defense against these compression adapted threats >

Priya: So, while this work is great for understanding how these attacks work, the next step in research should be developing detection and mitigation methods specifically for backdoor attacks under lossy compression >

Conclusion: Nadia: So we’re wrapping up on "Backdoor Attacks under Lossy Compression," which is essentially about how invisible triggers can get ruined when you compress them for storage and transmission, and this paper shows two ways to fix that failure.

Elias: Yeah, they're proposing Universal Attack Reactivation and Compression-Adapted Attack as solutions to keep those triggers alive after the lossy compression process.

Priya: From what I see, it’s about making sure the trigger survives the journey from being created in a clean image to being used by a deep learning model. It gets into how those subtle high-frequency details that make an attack stealthy are exactly what get distorted during JPEG or similar compressions.

Nadia: Exactly, and what’s interesting is that they show these methods work across different types of image compression, not just one specific codec.

Elias: They demonstrate effectiveness with all-to-one attack success rates over four different backbone models, which suggests these strategies aren't just theoretical ideas; they actually hold up in practice.

Priya: But we have to watch the caveats here. They acknowledge that this work focuses on building and revealing these attacks, not necessarily creating a perfect defense against them yet.

Nadia: That’s fair, and the paper points out that while CAA is pretty robust against things like Fine-Pruning defenses, they haven't solved every possible way to stop these compression adapted threats.

Elias: The implication for us is that if you’re building systems where data gets compressed before training, you can’t just assume a trigger will survive; you have to design the attack or the defense around that lossy step.

Priya: And what this means for people on the ground is that if we use these methods to understand vulnerabilities, we need more research into how to actually detect these compromised models after they've been stored and recompressed.

Nadia: Right, so while the attacks are getting better at surviving compression, the next big challenge is building a system that can reliably flag a model poisoned this way.

Elias: Exactly. So now we’ve seen how to build the problem under these real-world conditions; next time, we need to figure out how to stop it in the wild.

Qian Li, Yunuo Chen, ⨂Yuntian Chen

Shanghai Jiao Tong University · Eastern Institute of Technology

cs.CR, cs.CV

Submitted: 2026-03-14

Updated: 2026-11-04

Comments: Accepted to NeurIPS 2026

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 99/100

The gist: Real-world backdoor attacks often require poisoned datasets to be stored and transmitted before they are used to compromise deep learning systems.

Key concepts

Lossy Compression Failure
When an invisible backdoor is embedded in an image and then compressed (like JPEG), the resulting distorted data often loses the trigger's malicious information. This happens because lossy compression inevitably introduces errors, which can attenuate or completely destroy the subtle signals that activate the backdoor during inference.
Universal Attack Reactivation (UAR)
This strategy targets models that support pixel-level Region of Interest (ROI) coding. It works by measuring the difference between a clean and poisoned image to create ROI masks. These masks are used to selectively preserve or reactivate the trigger information that might have been degraded by compression.
Compression-Adapted Attack (CAA)
This is a new attack method that builds triggers intrinsically linked to the compression process. It uses customized ROI masks during encoding to strategically allocate bitrate, ensuring the malicious information is encoded in a way that survives both traditional codecs and learned image compression (LIC).

Terminology

Summary

Real-world backdoor attacks often require poisoned datasets to be stored and transmitted before they are used to compromise deep learning systems. p. 0

Where the paper says it

Backdoor Attacks under Lossy Compression: From Failure to Reactivation and Adaptation Qian Li1,2 Yunuo Chen1 Yuntian Chen2∗ 1Shanghai Jiao Tong University 2Eastern Institute of Technology Abstract Real-world backdoor attacks often require poisoned datasets to be stored and transmitted before they are used to compromise deep learning systems. In the era of big data, however, the inevitable use of lossy compression poses a fundamental challenge to invisible backdoor attacks.

The paper highlights the necessity of explicitly accounting for lossy compression in backdoor attacks. p. 2
Where the paper says it

Our contributions are summarized as follows: • We reveal that previously effective invisible triggers can become ineffective after lossy compression. This finding highlights the necessity of explicitly accounting for inevitable storage-and-transmission compression in data poisoning-based backdoor attacks.

The core problem addressed is the failure of invisible triggers under lossy compression. p. 0

Where the paper says it

Backdoor Attacks under Lossy Compression: From Failure to Reactivation and Adaptation Qian Li1,2 Yunuo Chen1 Yuntian Chen2∗ 1Shanghai Jiao Tong University 2Eastern Institute of Technology Abstract Real-world backdoor attacks often require poisoned datasets to be stored and transmitted before they are used to compromise deep learning systems. In the era of big data, however, the inevitable use of lossy compression poses a fundamental challenge to invisible backdoor attacks.

Invisible triggers embedded in RGB images may not survive the conversion from RGB data to compressed binary bitstreams and back to decompressed RGB images. p. 1

Where the paper says it

embedded in RGB images may not survive the conversion from RGB data to compressed binary bitstreams and back to decompressed RGB images. Formally, many invisible attacks construct poisoned samples by directly modifying RGB images, e. p. 1

Where the paper says it

embedded in RGB images may not survive the conversion from RGB data to compressed binary bitstreams and back to decompressed RGB images. Formally, many invisible attacks construct poisoned samples by directly modifying RGB images, e.

Formally, many invisible attacks construct poisoned samples by directly modifying RGB images, e.g., xp = x + δ ∈ R 3×H×W. Yet, once xp is lossily compressed into a binary bitstream yˆ ∈ 0, 1n, such as a JPEG file, there is no guarantee that the malicious trigger information is preserved in yˆ. p. 1
Where the paper says it

, xp = x + δ ∈ R 3×H×W . Yet, once xp is lossily compressed into a binary bitstream yˆ ∈ {0, 1} n, such as a JPEG file, there is no guarantee that the malicious trigger information is preserved in yˆ.

After decompression, the recovered poisoned image xˆp ∈ R 3×H×W may no longer contain an effective trigger. p. 1
Where the paper says it

Yet, once xp is lossily compressed into a binary bitstream yˆ ∈ {0, 1} n, such as a JPEG file, there is no guarantee that the malicious trigger information is preserved in yˆ. After decompression, the recovered poisoned image xˆp ∈ R 3×H×W may no longer contain an effective trigger.

Since lossy compression inevitably introduces distortion, i. p. 1
Where the paper says it

After decompression, the recovered poisoned image xˆp ∈ R 3×H×W may no longer contain an effective trigger. Since lossy compression inevitably introduces distortion, i.

e., xˆ ≠ x, it can attenuate or destroy trigger signals. p. 1
Where the paper says it

Since lossy compression inevitably introduces distortion, i. , xˆ ̸= x, it can attenuate or destroy trigger signals.

As a result, poisoned data that are effective in the raw RGB domain may lose their malicious functionality after storage and transmission, causing backdoor injection to fail. p. 1
Where the paper says it

, xˆ ̸= x, it can attenuate or destroy trigger signals. As a result, poisoned data that are effective in the raw RGB domain may lose their malicious functionality after storage and transmission, causing backdoor injection to fail.

This phenomenon affects a broad range of existing invisible attacks, including both invisible-trigger methods and clean-label attacks. p. 1
Where the paper says it

As a result, poisoned data that are effective in the raw RGB domain may lose their malicious functionality after storage and transmission, causing backdoor injection to fail. This phenomenon affects a broad range of existing invisible attacks, including both invisible-trigger methods and clean-label attacks [22, 23, 24, 25, 26, 27].

The authors propose two ROI-based strategies to counteract this failure. p. 1

Where the paper says it

This provides a natural interface for manipulating how trigger-related information is preserved or encoded during compression, without modifying the standard decompression process used by victims. Following this principle, we propose two ROI-based poisoning strategies tailored to inevitable lossy compression from failure to reactivation and adaptation.

Building on the region-of-interest (ROI) coding mechanism in image compression, they propose two poisoning strategies tailored to inevitable lossy compression. p. 0

Where the paper says it

This requires attackers to ensure that transmitted binary bitstreams preserve malicious trigger information, such that effective triggers can be induced after decompression. Building on the region-of-interest (ROI) coding mechanism in image compression, we propose two poisoning strategies tailored to inevitable lossy compression.

First, they introduce Universal Attack Reactivation, a general method that uses sample-specific ROI masks to reactivate trigger information in learned image compression (LIC). p. 0
Where the paper says it

Building on the region-of-interest (ROI) coding mechanism in image compression, we propose two poisoning strategies tailored to inevitable lossy compression. First, we introduce Universal Attack Reactivation, a general method that uses sample-specific ROI masks to reactivate trigger information in bitstreams for learned image compression (LIC).

Second, they present Compression-Adapted Attack (CAA), a new attack strategy that employs customized ROI masks to encode trigger information into bitstreams and applies to both traditional codecs and LIC. p. 0
Where the paper says it

First, we introduce Universal Attack Reactivation, a general method that uses sample-specific ROI masks to reactivate trigger information in bitstreams for learned image compression (LIC). Second, we present CompressionAdapted Attack, a new attack strategy that employs customized ROI masks to encode trigger information into bitstreams and applies to both traditional codecs and LIC.

Universal Attack Reactivation (UAR) focuses on preserving existing triggers. p. 1

Where the paper says it

The first, Universal Attack Reactivation, aims to reactivate existing invisible attacks that become ineffective after compression by preserving trigger-relevant information during learned image compression (LIC) [36, 37, 38, 39]. The second, Compression-Adapted Attack (CAA), goes one step further by designing triggers that are inherently adapted to the compression process, using customized ROI masks to encode malicious information into compressed bitstreams.

The first route, introduced in Sec. p. 4

Where the paper says it

We use this mechanism to design two complementary attack routes under lossy compression. The first route, introduced in Sec.

3.3, targets LIC models with pixel-level ROI support and reactivates existing invisible triggers by preserving trigger-relevant information. p. 4
Where the paper says it

The first route, introduced in Sec. 3, targets LIC models with pixel-level ROI support and reactivates existing invisible triggers by preserving trigger-relevant information.

Given a clean image x and its poisoned counterpart xp generated by an existing attack, a direct way to identify trigger-relevant regions is to measure their residuals as ROI masks: Mres = NormMeanx − xp, (3) where Mean(·) averages over RGB channels, and Norm(z) = z−min(z)max(z−min(z)) normalizes the mask values. p. 4
Where the paper says it

This strategy is particularly suitable for LICs with pixel-level ROI support, where fine-grained bitrate allocation can be controlled through sample-specific ROI masks. Given a clean image x and its poisoned counterpart xp generated by an existing attack, a direct way to identify trigger-relevant regions is to measure their residuals as ROI masks: Mres = NormMean|x − xp| , (3) where Mean(·) averages over RGB channels, and Norm(z) = z−min(z) max(z)−min(z) normalizes the mask values.

Compression-Adapted Attack (CAA) constructs triggers intrinsically coupled with the compression process. p. 5

Where the paper says it

Different from prior compression-aware attacks that treat compression artifacts as triggers or exploit distributional gaps between compressed poisoned samples and uncompressed benign samples, our goal is not to use compression merely as an additional trigger source. Instead, we design a new trigger that is intrinsically coupled with the lossy storage-and-transmission process.

The second route, introduced in Sec. p. 4

Where the paper says it

3, targets LIC models with pixel-level ROI support and reactivates existing invisible triggers by preserving trigger-relevant information. The second route, introduced in Sec.

3.4, targets both traditional codecs and LICs with region-level ROI support and constructs compression-adapted triggers by encoding malicious information through customized ROI masks. p. 4
Where the paper says it

The second route, introduced in Sec. 4, targets both traditional codecs and LICs with region-level ROI support and constructs compression-adapted triggers by encoding malicious information through customized ROI masks.

They use a customized ROI mask M∗ for poisoned samples and a uniform mask Muniform for benign samples, yielding xˆp = DecoderDeQ(Q(Encoder(xp,M∗))), (5) and xˆb = DecoderDeQ(Q(Encoder(xb,Muniform))). p. 5
Where the paper says it

Concretely, we use a customized ROI mask M∗ for poisoned samples and a uniform mask Muniform for benign samples. The decompressed poisoned and benign samples are respectively given by xˆp = DecoderDeQ(Q(Encoder(xp,M∗ ))) , (5) xˆb = DecoderDeQ(Q(Encoder(xb,Muniform))) .

(6) This encodes malicious information through bitrate allocation during compression. p. 5
Where the paper says it

The decompressed poisoned and benign samples are respectively given by xˆp = DecoderDeQ(Q(Encoder(xp,M∗ ))) , (5) xˆb = DecoderDeQ(Q(Encoder(xb,Muniform))) . (6) Unlike conventional invisible attacks that add perturbations in the RGB domain, CAA encodes malicious information through bitrate allocation during compression.

Experimental results demonstrate the effectiveness of both strategies. p. 0

Where the paper says it

Second, we present CompressionAdapted Attack, a new attack strategy that employs customized ROI masks to encode trigger information into bitstreams and applies to both traditional codecs and LIC. Extensive experiments demonstrate the effectiveness of both strategies.

Extensive experiments demonstrate the effectiveness of both strategies. p. 0

Where the paper says it

Second, we present CompressionAdapted Attack, a new attack strategy that employs customized ROI masks to encode trigger information into bitstreams and applies to both traditional codecs and LIC. Extensive experiments demonstrate the effectiveness of both strategies.

Table 1 reports all-to-one attack results for CAA, showing an ASR of over 99. p. 5
Where the paper says it

We first evaluate whether UAR can restore existing invisible backdoor attacks that fail after lossy compression. 1 ( ReA denotes the UAR-reactivated variant of a baseline attack) reports all-to-one attack results on different datasets and backbones.

9% on all four backbone models across all datasets. p. 7
Where the paper says it

Our method achieves ASR of over 99. 9% on all four backbone models across all datasets, while the BA remains comparable to the clean model.

Furthermore, CAA consistently achieves an ASR above 95% under JPEG re-compression when using a relatively sparse ROI mask. p. 8
Where the paper says it

In contrast, the performance of CAA depends on both the re-compression quality factor and the selected ROI mask. When the JPEG quality factor is set to the default or a higher value, and a relatively sparse ROI mask is adopted, CAA consistently achieves an ASR above 95%.

The paper concludes that the proposed methods are robust against various defenses. p. 7

Where the paper says it

Resistance to defense methods. We evaluate CAA method against the common defense methods, including Fine-Pruning [82], STRIP [83], and Gaussian noise and blur defenses.

The results of resisting Fine-Pruning are shown in Fig. p. 7

Where the paper says it

The introduction of defense methods is detailed in the supplementary material. • The results of resisting Fine-Pruning are shown in Fig.

5, which indicate that FinePruning cannot completely eliminate the backdoor implanted by our method. p. 7
Where the paper says it

• The results of resisting Fine-Pruning are shown in Fig. 5, which indicate that FinePruning cannot completely eliminate the backdoor implanted by our method.

The introduction of defense methods is detailed in the supplementary material. p. 7
Where the paper says it

We evaluate CAA method against the common defense methods, including Fine-Pruning [82], STRIP [83], and Gaussian noise and blur defenses. The introduction of defense methods is detailed in the supplementary material.

CAA shows stronger resistance than BadNet in all defenses. p. 7
Where the paper says it

Resistance to defense methods. We evaluate CAA method against the common defense methods, including Fine-Pruning [82], STRIP [83], and Gaussian noise and blur defenses.

Table 10 demonstrates that CAA remains effective under lossy compression while offering superior stealth. p. 8
Where the paper says it

Since prior invisible triggers fail under lossy compression, comparing stealth against those failed methods is meaningless; therefore we compare CAA with the visible attack BadNet [8]. 10 demonstrates that CAA remains effective under lossy compression while offering superior stealth.

The results of recompression show that CAA offers substantially stronger robustness to practical re-compression than non-adaptive invisible attacks. p. 8
Where the paper says it

When the JPEG quality factor is set to the default or a higher value, and a relatively sparse ROI mask is adopted, CAA consistently achieves an ASR above 95%. These results demonstrate that CAA offers substantially stronger robustness to practical re-compression than non-adaptive invisible attacks.

p. 8
Where the paper says it

When the JPEG quality factor is set to the default or a higher value, and a relatively sparse ROI mask is adopted, CAA consistently achieves an ASR above 95%. These results demonstrate that CAA offers substantially stronger robustness to practical re-compression than non-adaptive invisible attacks.

Limitations and future directions are acknowledged. p. 8

Where the paper says it

Additional results on complexity, defenses, and robustness are provided in Appendices A–F. 5 Conclusion and Limitations In this paper, we study backdoor attacks under inevitable lossy compression, a practical yet underexplored setting in real-world data storage and transmission pipelines.

A limitation of this work is that we focus on revealing and constructing compression-aware backdoor attacks, but do not propose a dedicated defense against such compression-adapted threats. p. 8

Where the paper says it

Experiments across multiple datasets, backbones, and codecs demonstrate the effectiveness of our methods. A limitation of this work is that we focus on revealing and constructing compression-aware backdoor attacks, but do not propose a dedicated defense against such compression-adapted threats.

p. 8
Where the paper says it

Experiments across multiple datasets, backbones, and codecs demonstrate the effectiveness of our methods. A limitation of this work is that we focus on revealing and constructing compression-aware backdoor attacks, but do not propose a dedicated defense against such compression-adapted threats.

Developing effective detection and mitigation methods for backdoor attacks under lossy compression remains an important direction for future research. p. 8
Where the paper says it

A limitation of this work is that we focus on revealing and constructing compression-aware backdoor attacks, but do not propose a dedicated defense against such compression-adapted threats. Developing effective detection and mitigation methods for backdoor attacks under lossy compression remains an important direction for future research.

p. 8
Where the paper says it

A limitation of this work is that we focus on revealing and constructing compression-aware backdoor attacks, but do not propose a dedicated defense against such compression-adapted threats. Developing effective detection and mitigation methods for backdoor attacks under lossy compression remains an important direction for future research.

The paper provides quantitative evidence of the proposed strategies. p. 1

Where the paper says it

This provides a natural interface for manipulating how trigger-related information is preserved or encoded during compression, without modifying the standard decompression process used by victims. Following this principle, we propose two ROI-based poisoning strategies tailored to inevitable lossy compression from failure to reactivation and adaptation.

Table 7 presents the performance on different codecs, showing that CAA outperforms the baseline methods in terms of both benign accuracy (BA) and attack success rate (ASR). p. 5

Where the paper says it

All images are resized to 64 × 64 and treated as raw images. We report Benign Accuracy (BA), Attack Success Rate (ASR) [75], and Attack Stealthiness (AS) to demonstrate the effectiveness of each attack method.

Table 8 reports ASRs of recompression with JPEG on GTSRB, where CAA achieves an ASR above 95% when the JPEG quality factor is set to the default or a higher value and a relatively sparse ROI mask is adopted. p. 8
Where the paper says it

In contrast, the performance of CAA depends on both the re-compression quality factor and the selected ROI mask. When the JPEG quality factor is set to the default or a higher value, and a relatively sparse ROI mask is adopted, CAA consistently achieves an ASR above 95%.

Table 10 shows that our method is imperceptible by both objective metrics and human perception. p. 8
Where the paper says it

Such frequency pattern serves as a trigger correlated with the target class, enabling a successful backdoor. 10 shows that our method is imperceptible by both objective metrics and human perception (human inspection test setting follows WaNet [24]).

The overall bit allocation is adjusted spatially rather than uniformly increased, suggesting improved attack effectiveness with little additional storage or transmission overhead. p. 8
Where the paper says it

Our method redistributes bits according to the ROI mask, assigning more bits to light (high-priority) regions while allocating fewer bits to dark regions. As a result, the overall bit allocation is adjusted spatially rather than uniformly increased.

The paper provides detailed visualizations of the attack mechanism. p. 0

Where the paper says it

Building on the region-of-interest (ROI) coding mechanism in image compression, we propose two poisoning strategies tailored to inevitable lossy compression. First, we introduce Universal Attack Reactivation, a general method that uses sample-specific ROI masks to reactivate trigger information in bitstreams for learned image compression (LIC).

Figure 7 is the visualization of CAA. p. 8

Where the paper says it

These results demonstrate that CAA offers substantially stronger robustness to practical re-compression than non-adaptive invisible attacks. Origin Compression-Adapted Residual Allocation Map (a) (b) (c) (d) Origin Poisoned by CAA Residual ROI mask Figure 7: Visualization of CAA.

Figure 8 further analyzes our CAA method-driven special frequency pattern by splitting each image into four regions according to the ROI mask and showing the DCT spectrum for each region. p. 8
Where the paper says it

All reported experiments of CAA use an 8×8 checkerboard. 8 further analyzes our CAA method driven special frequency pattern: we split each image into four regions according to the ROI mask and show the DCT spectrum for each region.

These poisoned images exhibit noticeably stronger high-frequency signals in the top-left and bottom-right regions than in the top-right and bottom-left. p. 8
Where the paper says it

8 further analyzes our CAA method driven special frequency pattern: we split each image into four regions according to the ROI mask and show the DCT spectrum for each region. The poisoned images exhibit noticeably stronger high-frequency signals in the top-left and bottom-right regions than in the top-right and bottom-left.

The paper compares CAA against existing compression-based methods. p. 2

Where the paper says it

While such methods improve robustness against compression, they typically require the attacker to participate in or control the training process of a specific classification model, for example by modifying the training objective. This differs from the dataset-only poisoning setting considered in this paper, where the attacker only releases a poisoned dataset and the victim trains a model using an arbitrary standard training pipeline.

Table 15, Table 16, Table 17, Table 18 and Table 19 present the complete comparison results of our CAA and the baseline methods on JPEG2000 [49], BPG[50], Cheng-2020 [81], and ELIC [59]. p. 6

Where the paper says it

01↑) 100. 00 Table 1: Comparison of methods across multiple datasets; additional results are in the Appendix.

p. 6
Where the paper says it

01↑) 100. 00 Table 1: Comparison of methods across multiple datasets; additional results are in the Appendix.

CAA outperforms the baseline methods in terms of both benign accuracy (BA) and attack success rate (ASR) across these codecs. p. 5
Where the paper says it

All images are resized to 64 × 64 and treated as raw images. We report Benign Accuracy (BA), Attack Success Rate (ASR) [75], and Attack Stealthiness (AS) to demonstrate the effectiveness of each attack method.

p. 5
Where the paper says it

All images are resized to 64 × 64 and treated as raw images. We report Benign Accuracy (BA), Attack Success Rate (ASR) [75], and Attack Stealthiness (AS) to demonstrate the effectiveness of each attack method.

Table 9 shows that our method does not noticeably increase file size or equivalently the resulting bit budget. p. 8
Where the paper says it

Consequently, the results in Tab. 9 show that our method does not noticeably increase file size, or equivalently, the resulting bit budget.

The paper summarizes the contribution of the two strategies. p. 8

Where the paper says it

5 Conclusion and Limitations In this paper, we study backdoor attacks under inevitable lossy compression, a practical yet underexplored setting in real-world data storage and transmission pipelines. We show that existing invisible backdoor attacks can become ineffective after compression, and propose two ROI-based strategies to address this challenge.

The contributions are summarized as follows:

• We reveal that previously effective invisible triggers can become ineffective after lossy compression.

• We propose Universal Attack Reactivation, a general ROI-based method for LICs that reactivates invisible triggers degraded by lossy compression. p. 2

Where the paper says it

This finding highlights the necessity of explicitly accounting for inevitable storage-and-transmission compression in data poisoning-based backdoor attacks. • We propose Universal Attack Reactivation, a general ROI-based method for LICs that reactivates invisible triggers degraded by lossy compression.

• We propose Compression-Adapted Attack (CAA), a new invisible backdoor attack for both LICs and traditional codecs. p. 2

Where the paper says it

Extensive experiments show that it significantly improves the attack success rates of prior invisible attacks under compression. • We propose Compression-Adapted Attack (CAA), a new invisible backdoor attack for both LICs and traditional codecs.

The paper provides detailed experimental settings.

We evaluate backdoor attack methods on GTSRB [72], CIFAR-10 [73], and CelebA [74]. p. 5

Where the paper says it

1 Experimental Setting Datasets and Metrics. We evaluate backdoor attack methods on GTSRB [72], CIFAR-10 [73], and CelebA [74].

We use ResNet [76], Pre-activation ResNet [77], MobileNetV2 [78], and SENet [79] as classifiers. p. 5
Where the paper says it

Classifiers. We use ResNet [76], Pre-activation ResNet [77], MobileNetV2 [78], and SENet [79].

We evaluate ROI-supported traditional codecs, including BPG [50] and JPEG 2000 [49], and ROI-integrated LIC models, including Mbt-2018 [80], Cheng-2020 [81], and ELIC [59]. p. 5
Where the paper says it

01, decayed by 10× every 100 epochs. We evaluate ROI-supported traditional codecs, including BPG [50] and JPEG 2000 [49], and ROI-integrated LIC models, including Mbt-2018 [80], Cheng-2020 [81], and ELIC [59].

We compare with three invisible attacks, WaNet [24], FTrojan [23], and BppAttack [22], two clean-label attacks, SAA [26] and Narcissus [27], and two compression-based methods, ECA [30] and CBA [2]. p. 5
Where the paper says it

We evaluate ROI-supported traditional codecs, including BPG [50] and JPEG 2000 [49], and ROI-integrated LIC models, including Mbt-2018 [80], Cheng-2020 [81], and ELIC [59]. We compare with three invisible attacks, WaNet [24], FTrojan [23], and BppAttack [22], two clean-label attacks, SAA [26] and Narcissus [27], and two compression-based methods, ECA [30] and CBA [2].

The paper discusses the relationship between triggers and frequency characteristics. p. 1

Where the paper says it

Right: Removing high-frequency components from poisoned test samples via Fourier Transform significantly reduces the ASR, suggesting that high-frequency information is critical for many invisible triggers. We observe that this failure is closely related to the frequency characteristics of invisible triggers.

We observe that this failure is closely related to the frequency characteristics of invisible triggers. p. 1

Where the paper says it

Right: Removing high-frequency components from poisoned test samples via Fourier Transform significantly reduces the ASR, suggesting that high-frequency information is critical for many invisible triggers. We observe that this failure is closely related to the frequency characteristics of invisible triggers.

Inspired by prior studies on frequency-domain backdoor behavior [28, 29], we find that many invisible triggers rely on subtle high-frequency components to achieve stealthiness. p. 1
Where the paper says it

We observe that this failure is closely related to the frequency characteristics of invisible triggers. Inspired by prior studies on frequency-domain backdoor behavior [28, 29], we find that many invisible triggers rely on subtle high-frequency components to achieve stealthiness.

These components are visually unobtrusive, but often play a crucial role in triggering the backdoor. p. 1
Where the paper says it

Inspired by prior studies on frequency-domain backdoor behavior [28, 29], we find that many invisible triggers rely on subtle high-frequency components to achieve stealthiness. These components are visually unobtrusive, but often play a crucial role in triggering the backdoor.

The paper details the robustness of CAA against specific adversarial defenses. p. 2

Where the paper says it

While such methods improve robustness against compression, they typically require the attacker to participate in or control the training process of a specific classification model, for example by modifying the training objective. This differs from the dataset-only poisoning setting considered in this paper, where the attacker only releases a poisoned dataset and the victim trains a model using an arbitrary standard training pipeline.

In Tab. 10, we demonstrate that CAA remains effective under lossy compression while offering superior stealth compared to the visible attack BadNet [8]. p. 8

Where the paper says it

Since prior invisible triggers fail under lossy compression, comparing stealth against those failed methods is meaningless; therefore we compare CAA with the visible attack BadNet [8]. 10 demonstrates that CAA remains effective under lossy compression while offering superior stealth.

The results in Fig. p. 7
Where the paper says it

The introduction of defense methods is detailed in the supplementary material. • The results of resisting Fine-Pruning are shown in Fig.

6 indicate that the entropy ranges of the clean model and the backdoored model trained on our poisoned samples are very similar, suggesting resistance to STRIP in compression scenarios. p. 7
Where the paper says it

6 indicate that the entropy ranges of the clean model and the backdoored model trained on our poisoned samples, are very similar. Our method is resistant to STRIP in compression scenarios.

After applying Gaussian blur (kernel:3) to decompressed test samples, we achieved a BA of 91. p. 7
Where the paper says it

Our method is resistant to STRIP in compression scenarios. • After applying Gaussian blur (kernel:3) to decompressed test samples, we achieved a BA of 91.

62% and an ASR of 98. p. 1
Where the paper says it

Left: Higher compression rates introduce stronger distortion and substantially hinder backdoor injection, indicating that lossy compression can severely damage invisible triggers. Right: Removing high-frequency components from poisoned test samples via Fourier Transform significantly reduces the ASR, suggesting that high-frequency information is critical for many invisible triggers.

72%.

The paper concludes with a statement on the generalizability of CAA. p. 7

Where the paper says it

Our method consistently demonstrates minimal BA loss and high ASR close to 100% across different codecs. This performance underscores the broad applicability and generalizability of CAA.

Our method consistently demonstrates minimal BA loss and high ASR close to 100% across different codecs. p. 7

Where the paper says it

ResNet18 [76] is employed as the classification model. Our method consistently demonstrates minimal BA loss and high ASR close to 100% across different codecs.

This performance underscores the broad applicability and generalizability of CAA. p. 7
Where the paper says it

Our method consistently demonstrates minimal BA loss and high ASR close to 100% across different codecs. This performance underscores the broad applicability and generalizability of CAA.

The results in Tab. p. 8
Where the paper says it

As a result, the overall bit allocation is adjusted spatially rather than uniformly increased. Consequently, the results in Tab.

p. 8
Where the paper says it

As a result, the overall bit allocation is adjusted spatially rather than uniformly increased. Consequently, the results in Tab.

p. 8
Where the paper says it

As a result, the overall bit allocation is adjusted spatially rather than uniformly increased. Consequently, the results in Tab.

9 show that our method does not noticeably increase file size or equivalently the resulting bit budget. p. 8
Where the paper says it

Consequently, the results in Tab. 9 show that our method does not noticeably increase file size, or equivalently, the resulting bit budget.

The paper acknowledges its limitations. p. 8

Where the paper says it

Additional results on complexity, defenses, and robustness are provided in Appendices A–F. 5 Conclusion and Limitations In this paper, we study backdoor attacks under inevitable lossy compression, a practical yet underexplored setting in real-world data storage and transmission pipelines.

A limitation of this work is that we focus on revealing and constructing compression-aware backdoor attacks, but do not propose a dedicated defense against such compression-adapted threats. Developing effective detection and mitigation methods for backdoor attacks under lossy compression remains an important direction for future research.

The paper provides quantitative evaluation of the reactivation method. p. 8

Where the paper says it

000 - BadNet 21. 4 Table 10: Quantitative evaluation of stealthiness.

Table 4 shows that UAR mainly preserves and reactivates existing trigger information instead of creating a new mask-only trigger. p. 6

Where the paper says it

Removing both the trigger and Mfreq further reduces ASR to nearly zero. Thus, UAR mainly preserves and reactivates existing trigger information instead of creating a new mask-only trigger.

Ablation studies of the ROI mask verify whether UAR reactivates original triggers rather than introducing a new ROI-induced shortcut. p. 6
Where the paper says it

Ablation studies of the ROI mask. 4 verifies whether UAR reactivates original triggers rather than introducing a new ROI-induced shortcut.

The paper provides quantitative evaluation of the adaptation method. p. 8

Where the paper says it

000 - BadNet 21. 4 Table 10: Quantitative evaluation of stealthiness.

Table 11 shows that our CAA mislabels nearly 100% of poisoned images with only a 0. p. 7

Where the paper says it

Our method outperforms alternatives across three datasets. Our CAA mislabels nearly 100% of poisoned images with only a 0.

06% loss in BA. p. 7
Where the paper says it

Our CAA mislabels nearly 100% of poisoned images with only a 0. 06% loss in BA.

The results in Table 7 demonstrate that CAA consistently achieves minimal BA loss and high ASR close to 100% across different codecs. p. 7
Where the paper says it

Our method consistently demonstrates minimal BA loss and high ASR close to 100% across different codecs. This performance underscores the broad applicability and generalizability of CAA.

The paper provides quantitative evaluation of the compression-adapted attack. p. 7

Where the paper says it

We visualize original images, compressed attack samples, reactivated samples, and the corresponding ROI masks. 3 Evaluation of Compression-Adapted Attack All-to-one attack.

Table 15, Table 16, Table 17, Table 18 and Table 19 present the complete comparison results of our CAA and the baseline methods on JPEG2000 [49], BPG[50], Cheng-2020 [81], and ELIC [59]. CAA outperforms the baseline methods in terms of both benign accuracy (BA) and attack success rate (ASR) across these codecs. The performance on different compression models consistently demonstrates minimal BA loss and high ASR close to 100%. p. 7

Where the paper says it

Our method consistently demonstrates minimal BA loss and high ASR close to 100% across different codecs. This performance underscores the broad applicability and generalizability of CAA.

The paper provides a comprehensive comparison across multiple backbones. p. 6

Where the paper says it

01↑) 100. 00 Table 1: Comparison of methods across multiple datasets; additional results are in the Appendix.

Table 16 shows that CAA outperforms the baseline methods in terms of both benign accuracy (BA) and attack success rate (ASR) across different backbones. p. 5

Where the paper says it

All images are resized to 64 × 64 and treated as raw images. We report Benign Accuracy (BA), Attack Success Rate (ASR) [75], and Attack Stealthiness (AS) to demonstrate the effectiveness of each attack method.

The results in Table 15 show that CAA outperforms the baseline methods on JPEG [5]. p. 6
Where the paper says it

01↑) 100. 00 Table 1: Comparison of methods across multiple datasets; additional results are in the Appendix.

The paper provides a summary of the findings. p. 1

Where the paper says it

ROI coding allows codecs to assign different bitrates to different image regions, thereby controlling the spatial distribution of compression distortion. This provides a natural interface for manipulating how trigger-related information is preserved or encoded during compression, without modifying the standard decompression process used by victims.

In this paper, we study backdoor attacks under inevitable lossy compression, a practical yet underexplored setting in real-world data storage and transmission pipelines. p. 8

Where the paper says it

Additional results on complexity, defenses, and robustness are provided in Appendices A–F. 5 Conclusion and Limitations In this paper, we study backdoor attacks under inevitable lossy compression, a practical yet underexplored setting in real-world data storage and transmission pipelines.

We show that existing invisible backdoor attacks can become ineffective after compression and propose two ROI-based strategies to address this challenge. p. 8
Where the paper says it

5 Conclusion and Limitations In this paper, we study backdoor attacks under inevitable lossy compression, a practical yet underexplored setting in real-world data storage and transmission pipelines. We show that existing invisible backdoor attacks can become ineffective after compression, and propose two ROI-based strategies to address this challenge.

Experiments across multiple datasets, backbones, and codecs demonstrate the effectiveness of our methods. p. 8
Where the paper says it

We show that existing invisible backdoor attacks can become ineffective after compression, and propose two ROI-based strategies to address this challenge. Experiments across multiple datasets, backbones, and codecs demonstrate the effectiveness of our methods.

The paper provides a detailed overview of the pipeline. p. 0

Where the paper says it

Prior compression-based attacks typically exploit compression artifacts as certain triggers to distinguish poisoned RGB samples from uncompressed benign ones, rather than addressing whether malicious information can survive a shared lossy storage-and-transmission pipeline. In this paper, we highlight the necessity of explicitly accounting for lossy compression in backdoor attacks.

Figure 2 illustrates the dataset-only poisoning pipeline considered in this paper. p. 3

Where the paper says it

1 Threat Model and Problem Setup under Lossy Compression Fig. 2 illustrates the dataset-only poisoning pipeline considered in this paper.

Column 1 shows the real-world dataset-only poisoning pipeline, where data are stored and transmitted as compressed bitstreams before being used for victim-side training. p. 3
Where the paper says it

Red borders denote poisoned samples, and green borders denote benign samples. Column 1 shows the real-world dataset-only poisoning pipeline, where data are stored and transmitted as compressed bitstreams before victimside training.

Column 2 illustrates that previous invisible attacks may fail after lossy compression because trigger information is destroyed. p. 3
Where the paper says it

Column 1 shows the real-world dataset-only poisoning pipeline, where data are stored and transmitted as compressed bitstreams before victimside training. Column 2 illustrates that previous invisible attacks may fail after lossy compression because trigger information is destroyed.

Columns 3-4 present our two ROI-based strategies: Universal Attack Reactivation, which reactivates ineffective invisible attacks by preserving trigger-relevant information, and Compression-Adapted Attack, which encodes malicious information into bitstreams to construct compression-adapted triggers. p. 3
Where the paper says it

Column 2 illustrates that previous invisible attacks may fail after lossy compression because trigger information is destroyed. Columns 3-4 present our two ROI-based strategies: Universal Attack Reactivation, which reactivates ineffective invisible attacks by preserving triggerrelevant information, and Compression-Adapted Attack, which encodes malicious information into bitstreams to construct compression-adapted triggers for both traditional codecs and LICs.

The paper provides a detailed overview of the ROI coding mechanism. p. 1

Where the paper says it

ROI coding allows codecs to assign different bitrates to different image regions, thereby controlling the spatial distribution of compression distortion. This provides a natural interface for manipulating how trigger-related information is preserved or encoded during compression, without modifying the standard decompression process used by victims.

ROI coding allows codecs to assign different bitrates to different image regions, thereby controlling the spatial distribution of compression distortion. p. 1

Where the paper says it

We achieve this by leveraging the Region of Interest (ROI) coding mechanism in image compression [31, 32, 33, 34, 35]. ROI coding allows codecs to assign different bitrates to different image regions, thereby controlling the spatial distribution of compression distortion.

This interface is well suited to our goal, allowing the attacker to guide the standard compression process itself. p. 4
Where the paper says it

This interface is well suited to our goal. Instead of introducing additional backdoor modules or requiring a modified decoder, ROI masks allow the attacker to guide the standard compression process itself.

The ROI-supported encoder takes an additional mask M = [mij] ∈ R H×W as input: Compression: yˆ = Q(Encoder(x,M)). p. 4
Where the paper says it

69, 32, 33]. An ROI-supported encoder takes an additional mask M = [mij ] ∈ R H×W as input: Compression : yˆ = Q(Encoder(x,M)).

(2).

The paper discusses the relationship between compression and trigger information. p. 0

Where the paper says it

In this paper, we highlight the necessity of explicitly accounting for lossy compression in backdoor attacks. This requires attackers to ensure that transmitted binary bitstreams preserve malicious trigger information, such that effective triggers can be induced after decompression.

Stronger compression substantially reduces the attack success rate, and explicitly removing high-frequency components from poisoned test samples also leads to a significant ASR drop. p. 1

Where the paper says it

As shown in Fig. 1, stronger compression substantially reduces the attack success rate, and explicitly removing high-frequency components from poisoned test samples also leads to a significant ASR drop.

These observations suggest that invisible triggers achieve stealth by hiding in compression-sensitive high-frequency information, while lossy compression tends to suppress or distort exactly such information. p. 1
Where the paper says it

1, stronger compression substantially reduces the attack success rate, and explicitly removing high-frequency components from poisoned test samples also leads to a significant ASR drop. These observations suggest that invisible triggers achieve stealth by hiding in compression-sensitive high-frequency information, while lossy compression tends to suppress or distort exactly such information.

The paper provides a final statement on the study's scope.

We consider storage and transmission with publicly available codec implementations or specifications, so the attack is evaluated under realistic compression pipelines. p. 3

Where the paper says it

The attacker aims to release a poisoned dataset whose decompressed samples can successfully implant a backdoor into a victim-trained model, while maintaining normal accuracy on decompressed benign samples. We consider storage and transmission with publicly available codec implementations or specifications, so the attack is evaluated under realistic compression pipelines.

Our goal is therefore not merely to use compression artifacts as another form of trigger. p. 1
Where the paper says it

In contrast, they do not explicitly address whether malicious information can survive a shared lossy storage-and-transmission pipeline, where both benign and poisoned data are compressed into binary bitstreams. Our goal is therefore not merely to use compression artifacts as another form of trigger.

Instead, we study backdoor attacks under inevitable lossy compression, where the attacker must ensure that the transmitted bitstream recovers or induces malicious trigger information after decompression. p. 1
Where the paper says it

Our goal is therefore not merely to use compression artifacts as another form of trigger. Instead, we study backdoor attacks under inevitable lossy compression, where the attacker must ensure that the transmitted bitstream recovers or induces malicious trigger information after decompression.

The paper is a preprint arXiv:2603. p. 0
Where the paper says it

This introduces a fundamental challenge to invisible backdoor attacks: triggers ∗Corresponding Author Preprint. arXiv:2603.

13864v2 [cs. p. 0
Where the paper says it

arXiv:2603. 13864v2 [cs.

CR] 15 May 2026. p. 0
Where the paper says it

13864v2 [cs. CR] 15 May 2026

The paper details the specific attack strategies. p. 0

Where the paper says it

Building on the region-of-interest (ROI) coding mechanism in image compression, we propose two poisoning strategies tailored to inevitable lossy compression. First, we introduce Universal Attack Reactivation, a general method that uses sample-specific ROI masks to reactivate trigger information in bitstreams for learned image compression (LIC).

To address this challenge, we explicitly incorporate lossy compression into the design of backdoor attacks. p. 1

Where the paper says it

Instead, we study backdoor attacks under inevitable lossy compression, where the attacker must ensure that the transmitted bitstream recovers or induces malicious trigger information after decompression. To address this challenge, we explicitly incorporate lossy compression into the design of backdoor attacks.

We use this mechanism to design two complementary attack routes under lossy compression. p. 4
Where the paper says it

Instead of introducing additional backdoor modules or requiring a modified decoder, ROI masks allow the attacker to guide the standard compression process itself. We use this mechanism to design two complementary attack routes under lossy compression.

The first route targets LIC models with pixel-level ROI support and reactivates existing invisible triggers by preserving trigger-relevant information. p. 4
Where the paper says it

The first route, introduced in Sec. 3, targets LIC models with pixel-level ROI support and reactivates existing invisible triggers by preserving trigger-relevant information.

The second route targets both traditional codecs and LICs with region-level ROI support and constructs compression-adapted triggers by encoding malicious information through customized ROI masks. p. 4
Where the paper says it

The second route, introduced in Sec. 4, targets both traditional codecs and LICs with region-level ROI support and constructs compression-adapted triggers by encoding malicious information through customized ROI masks.

The paper provides a final statement on resilience. p. 7

Where the paper says it

86% and ASR reached 100%. These results demonstrate our method’s resilience against such defenses.

CAA offers substantially stronger robustness to practical re-compression than non-adaptive invisible attacks. The results in Tab. 8 report the ASR under JPEG re-compression, where CAA consistently achieves an ASR above 95% when the JPEG quality factor is set to the default or a higher value and a relatively sparse ROI mask is adopted. p. 8

Where the paper says it

In contrast, the performance of CAA depends on both the re-compression quality factor and the selected ROI mask. When the JPEG quality factor is set to the default or a higher value, and a relatively sparse ROI mask is adopted, CAA consistently achieves an ASR above 95%.

The performance of CAA depends on both the re-compression quality factor and the selected ROI mask. p. 8
Where the paper says it

Since FTrojan is not designed to be compressionadaptive, its trigger remains highly vulnerable to re-compression, even when it is injected before the initial compression stage. In contrast, the performance of CAA depends on both the re-compression quality factor and the selected ROI mask.

The results in Tab. 8 report the ASR under JPEG re-compression, where FTrojanReA almost completely fails. p. 8
Where the paper says it

8 reports the ASR under JPEG re-compression. As expected, FTrojanReA almost completely fails.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Universal Attack Reactivation (UAR) improves attack success rates of prior invisible attacks under compression by preserving trigger-relevant information during learned image compression (LIC). This allows for the reactivation of triggers that otherwise become ineffective after lossy compression.

  2. Compression-Adapted Attack (CAA) constructs triggers that are intrinsically coupled with the lossy storage-and-transmission process by using ROI masks to encode malicious information into compressed bitstreams. This induces a distinctive yet visually imperceptible frequency-distribution pattern in decompressed poisoned images which serves as a compression-adapted trigger.

Abstract

Real-world backdoor attacks often require poisoned datasets to be stored and transmitted before they are used to compromise deep learning systems. In the era of big data, however, the inevitable use of lossy compression poses a fundamental challenge to invisible backdoor attacks. We observe that triggers embedded in RGB images can become ineffective once the images are lossily compressed into binary bitstreams, such as JPEG files, for storage and transmission. Consequently, poisoned data may lose their malicious functionality after compression, causing backdoor injection to fail. Prior compression-based attacks typically exploit compression artifacts as certain triggers to distinguish poisoned RGB samples from uncompressed benign ones, rather than addressing whether malicious information can survive a shared lossy storage-and-transmission pipeline. In this paper, we highlight the necessity of explicitly accounting for lossy compression in backdoor attacks. This requires attackers to ensure that transmitted binary bitstreams preserve malicious trigger information, such that effective triggers can be induced after decompression. Building on the region-of-interest (ROI) coding mechanism in image compression, we propose two poisoning strategies tailored to inevitable lossy compression. First, we introduce Universal Attack Reactivation, a general method that uses sample-specific ROI masks to reactivate trigger information in bitstreams for learned image compression (LIC). Second, we present Compression-Adapted Attack, a new attack strategy that employs customized ROI masks to encode trigger information into bitstreams and applies to both traditional codecs and LIC. Extensive experiments demonstrate the effectiveness of both strategies.

Sources

Related papers