Evaluating Multimodal Steganalysis for Split-Payload Audiovisual Steganography

arXiv:2606.08726 · cs.CR · Submitted 2026-06-07 · 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: We're looking at the paper: "Evaluating Multimodal Steganalysis for Split-Payload Audiovisual Steganography".

Elias: Here's the gist of the paper. The aim of steganography is to hide secret information inside ordinary media so that the existence of communication is hidden rather than encrypted.

Nadia: This paper evaluates whether such split-payload audiovisual steganography can help evade unimodal and multimodal steganalysis under synchronized and asynchronous embedding settings.

Elias: The core investigation The central question is whether this split-payload design is harder to detect than conventional single-medium embedding, and whether a multimodal detector can recover a reliable signal when neither track alone provides one.

Nadia: The study constructs samples under several related embedding modes and evaluates all of them against a set of unimodal and multimodal detectors to map out the conditions under which the approach succeeds or fails.

Title and authors: Nadia: So we're looking at this paper today, "Evaluating Multimodal Steganalysis for Split-Payload Audiovisual Steganography," and it deals with splitting a hidden message between audio and video streams. That means instead of hiding everything in one medium, you spread the secret info across both video and audio.

Elias: Exactly. The authors are trying to figure out if this split payload design can help evade detection methods that look at either the video or the audio separately, or if a system that looks at both can still pull it out.

Priya: So basically, they’re testing if hiding things in two different formats makes it harder for steganalysis tools to catch the secret information. I wonder what kind of media this is being done on though.

Nadia: It's audiovisual context, so we're talking about video and audio streams together, which opens up a new way to hide data by dividing the load between those two modes. This paper asks if that specific method can dodge both single-mode and multi-mode detectors under different embedding settings like synchronized or asynchronous ones.

Elias: The core idea is that you calibrate the embedding rates so they fall below what a single detector would flag for either track alone, which is interesting because it shows the benefit of this hiding mechanism.

The paper's summary: Nadia: The paper sets up a binary hypothesis test: H0, where nothing is hidden in the cover material, and H1, where there is a secret message inside. They create these audiovisual samples where the hidden message is split between the audio and video tracks.

Elias: Then they test how well different detectors can identify it. They find that single-mode detectors perform close to random guessing, which shows how helpful this hiding mechanism is for evading detection on a per-track basis.

Priya: But then they look at the multimodal model initially, and it seems more effective than the single modes, but they find that this improvement mostly comes from the video stream rather than a real detection of an artifact.

Nadia: The main finding here is that even though a multimodal detector seems better at first glance, further checks reveal that it doesn't really detect the message in this split-payload condition.

Elias: They found that when they mask each modality and shuffle the audio video pairs, the cross-attention detector doesn't detect anything hidden in the split payload condition.

Priya: So what does this mean for us? It suggests that embedding at low rates across multiple modalities prevents a multimodal detector from learning a combined artifact.

The paper's improvements: Nadia: The authors suggest several ways to improve this approach, specifically by focusing on how they set up the split-payload embedding. They introduce improvements like splitting the payload across audio and video tracks at individually sub-threshold rates of zero point two five bpp for video and zero point zero eight bps for audio.

Elias: That is a key improvement because it directly addresses the problem where the unimodal baselines are performing at chance level, which is shown by a probability of five zero on those tests.

Priya: And this calibration seems important because it calibrates the perceptual distortion in each modality to be comparable, so you're not just hiding data in one stream while messing up the other.

Nadia: Plus, they suggest incorporating a shared temporal module using a bidirectional LSTM over K slices to ensure that any performance differences reflect a fusion strategy rather than just the model’s capacity for modeling time dependencies.

Elias: That’s important because it helps detect synchronization dependencies in synchronized embedding modes, which is what happens when both video and audio embedders share a common per-slice seed.

Priya: And for the video side, they refined feature extraction by using an SRNet-style encoder with absolute-value activation to extract residuals that are symmetric around plus or minus one LSB modifications, which reduces the risk of any observed multimodal advantage stemming from trivially detectable visual artifacts.

Conclusion: Nadia: So to wrap up on this paper, we confirm the correctness of this split-payload evasion strategy. The findings emphasize how important it is to do ablation analysis when you're doing multimodal steganalysis.

Elias: The lack of audio modality impact and the lack of sensitivity to shuffling pairs across samples serve as strong evidence for how difficult it is to detect crossmodal embedding in this way.

Priya: And relying solely on the video branch serves as another piece of evidence that makes detecting crossmodal embedding quite tough.

Nadia: The key implication here is that you can hide information by embedding at low rates across multiple modalities, and this prevents a multimodal detector from learning the combined artifact.

Elias: We should probably extend this experiment to multi-speaker datasets and explore if we can detect the crossmodal signal at higher payload rates.

Prateek Paudel, Nitin Jha, Abhishek Parakh

Kennesaw State University

cs.CR

Submitted: 2026-06-07

Updated: 2026-06-07

DOI: 10.1109/IMNS67862.2026.11655295

License: http://creativecommons.org/publicdomain/zero/1.0/

Importance score: 74/100

The gist: The aim of steganography is to hide secret information inside ordinary media so that the existence of communication is hidden rather than encrypted.

Key concepts

Split-Payload Steganography
This technique hides secret information by dividing the hidden data across different media types, like video and audio. Instead of embedding everything in one file, parts go into the video and other parts into the audio. The goal is to make it harder for detectors that look at only one type of media to find the hidden message.
Multimodal Steganalysis
This refers to detection methods designed to analyze multiple types of media simultaneously, such as looking at both a video and an audio file together. These detectors try to find statistical patterns or artifacts that appear when data is hidden in different formats at the same time. The paper tests if these combined detectors can reliably find hidden messages.
Cross-Attention Transformer
This is a specific type of deep learning architecture used as a detector. It allows the model to look at information from both modalities (video and audio) simultaneously and weigh their importance relative to each other. The paper uses this structure to see if it can successfully identify the hidden signal in split-payload scenarios.
Embedding Modes
These describe how the secret data is physically placed inside the video and audio files. The study tests different ways, including 'sync mode' where both tracks share a common seed, and 'asynchronous mode' where they use independent seeds. Understanding these modes helps determine if the hiding method can evade detection under different embedding conditions.

Terminology

Summary

The aim of steganography is to hide secret information inside ordinary media so that the existence of communication is hidden rather than encrypted. p. 0

Where the paper says it

Evaluating Multimodal Steganalysis for Split-Payload Audiovisual Steganography Prateek Paudel†*, Nitin Jha† , Abhishek Parakh† , †Kennesaw State University, Marietta, GA, USA *Corresponding Author: ppaudel@students. edu Abstract—The aim of steganography is to hide secret information inside ordinary media so that the existence of communication is hidden rather than encrypted.

This paper evaluates whether such split-payload audiovisual steganography can help evade unimodal and multimodal steganalysis under synchronized and asynchronous embedding settings. p. 0
Where the paper says it

In audiovisual context, the availability of audio and video streams creates an opportunity to split a payload across these two modes thus, reducing the embedding burden on any single carrier. This paper evaluates whether such split-payload audiovisual steganography can help evade unimodal and multimodal steganalysis under synchronized and asynchronous embedding settings.

The core investigation

The central question is whether this split-payload design is harder to detect than conventional single-medium embedding, and whether a multimodal detector can recover a reliable signal when neither track alone provides one. The study constructs samples under several related embedding modes and evaluates all of them against a set of unimodal and multimodal detectors to map out the conditions under which the approach succeeds or fails. p. 0

Where the paper says it

The central question is whether this split-payload design is harder to detect than conventional single-medium embedding, and whether a multimodal detector can recover a reliable signal when neither track alone provides one. We construct samples under several related embedding modes and evaluate all of them against a set of unimodal and multimodal detectors to map out the conditions under which the approach succeeds or fails.

Experimental Setup Details

The problem setup frames detection as a binary hypothesis test: H0: (V, A) ∼ Pcover, H1: (V, A) ∼ Pstego. Both modalities are divided into K contiguous temporal slices before embedding. p. 2

Where the paper says it

Whether a multimodal detector can exploit it when neither marginal signal is individually reliable is the question the experiments address. Temporal Slicing and Synchronization Both modalities are divided into K contiguous temporal slices before embedding.

The master seed produces a per-slice seed vector s = (s0,. p. 2
Where the paper says it

Temporal Slicing and Synchronization Both modalities are divided into K contiguous temporal slices before embedding. A clip-level master seed produces a per-slice seed vector s = (s0, .

.., sK−1) and a payload-rate modulation vector λ. p. 2
Where the paper says it

A clip-level master seed produces a per-slice seed vector s = (s0, . , sK−1) and a payload-rate modulation vector λ, drawn uniformly and clipped to a stable range.

In sync mode, both the video and audio embedders share a common per-slice seed, introducing a statistical dependency between the modification patterns of the two tracks. p. 2
Where the paper says it

Each track is embedded at a rate calibrated to fall below the detection threshold of its corresponding unimodal classifier. In the synchronized variant, both embedders share a common per-slice seed, introducing a statistical dependency between the modification patterns of the two tracks: I(ϕ v k ; ϕ a k | sk) > 0 (sync).

Embedding Pipeline Construction

The embedding pipeline involves specific algorithms for each modality.

  1. Video embedding uses the WOW algorithm [6], a content-adaptive scheme that concentrates modifications in textured regions. p. 2

    Where the paper says it

    Embedding Pipeline a) Video. : Video frames are embedded using the WOW algorithm [6], a content-adaptive scheme that concentrates modifications in textured regions.

  2. Audio embedding utilizes a syndrome-trellis code (STC) with energy-based per-sample costs. p. 2

    Where the paper says it

    WOW is chosen because its content-adaptive placement raises the bar for unimodal frame-level detectors, reducing the risk that any observed multimodal advantage stems from trivially detectable visual artifacts. : Audio samples are embedded using a syndrome-trellis code (STC) with energy-based per-sample costs: ρ a n = 1 RMS(n) + ϵ , RMS(n) = vuut 1 W W X−1 l=0 a 2 n−l .

  3. A distortion equalization step calibrates the audio payload rate β0 so that perceptual distortion in each modality is comparable. p. 2

    Where the paper says it

    c) Distortion equalization. : To prevent the detection problem from reducing to unimodal analysis of a single heavily loaded track, the audio payload rate β0 is calibrated via binary search so that perceptual distortion in each modality is comparable: |∆v(α0) − ∆a(β0)| < τ, (6) where ∆v = 1 − SSIM is video distortion and ∆a is PESQnormalized audio distortion.

Detector Architectures and Findings

The multimodal detector primarily consists of three bidirectional cross-attention transformer layers. The results show that the cross-attention detector demonstrates 93. p. 3

Where the paper says it

RESULTS We consider two major questions here. First, is the proposed cross-attention detector capable of detecting steganographic artifacts in the split-payload setting?

62% of test accuracy (PE = 0. p. 3
Where the paper says it

Crossattention detector demonstrates 93. 62% of test accuracy (PE = 0.

064) on synchronized split-payload embedding. p. 3
Where the paper says it

62% of test accuracy (PE = 0. 064) on synchronized split-payload embedding, while both

However, further analysis reveals that this result cannot be explained by detection of steganographic artifacts. p. 4
Where the paper says it

62% test accuracy. However, further analysis showed that this result cannot be explained by detection of steganographic artifacts.

Modality Contribution Analysis

A combination of masking of each modality and shuffling audio-video pairs leads to the conclusion that cross-attention detector does not detect the hidden message in the splitpayload condition. Specifically, removing the video stream drops test accuracy to chance level PE = 0. p. 4

Where the paper says it

For async mode, the effect is exactly zero. Second, removal of the video stream drops test accuracy to chance level PE = 0.

50. This suggests that the approach itself appears to be correct as embedding at low rates in multiple modalities prevents multimodal detector from learning the combined embedding artifact. p. 4
Where the paper says it

First, the approach itself appears to be correct: embedding at low rates in multiple modalities (0. 08 bps) prevents multimodal detector from learning the combined embedding artifact, even if it uses explicit cross-modality attention.

Final Conclusions and Implications

The findings confirm the correctness of the splitpayload evasion strategy, while emphasizing the importance of ablation-based analysis in multimodal steganalysis. The lack of audio modality impact, lack of sensitivity to shuffling of pairs across samples, and reliance solely on video branch serve as strong evidences for difficulty of crossmodal embedding detection. p. 5

Where the paper says it

These findings confirm the correctness of the splitpayload evasion strategy, while emphasizing the importance of ablation-based analysis in multimodal steganalysis experiments. Lack of audio modality impact, lack of sensitivity to shuffling of pairs across samples, and reliance solely on video branch, serve as strong evidences for difficulty of crossmodal embedding detection.

Further research should extend the experiment to multi-speaker datasets and explore the possibility of detecting cross-modal signal at higher payload rates. p. 5
Where the paper says it

Lack of audio modality impact, lack of sensitivity to shuffling of pairs across samples, and reliance solely on video branch, serve as strong evidences for difficulty of crossmodal embedding detection. Further research should extend the experiment to multi-speaker datasets and explore the possibility of detecting cross-modal signal at higher payload rates.

References

[1] SM Masud Karim, Md Saifur Rahman, and Md Ismail Hossain. “A new approach for LSB based image steganography using secret key”. p. 5

Where the paper says it

REFERENCES [1] SM Masud Karim, Md Saifur Rahman, and Md Ismail Hossain. “A new approach for LSB based image steganography using secret key”.

In: 14th international conference on computer and information technology (ICCIT 2011). p. 5
Where the paper says it

“A new approach for LSB based image steganography using secret key”. In: 14th international conference on computer and information technology (ICCIT 2011).

IEEE. 2011, pp. 286–291.

[3] Meike Helena Kombrink, Zeno Jean Marius Hubert Geradts, and Marcel Worring. p. 5

Where the paper says it

1 (2014), p. p. 5

Where the paper says it

“Edge-based image steganography”. In: EURASIP Journal on Information Security 2014.

[3] Meike Helena Kombrink, Zeno Jean Marius Hubert Geradts, and Marcel Worring.

“Image steganography approaches and their detection strategies: A survey”. p. 5
Where the paper says it

[3] Meike Helena Kombrink, Zeno Jean Marius Hubert Geradts, and Marcel Worring. “Image steganography approaches and their detection strategies: A survey”.

In: ACM Computing Surveys 57. p. 5
Where the paper says it

“Image steganography approaches and their detection strategies: A survey”. In: ACM Computing Surveys 57.

2 (2024), pp. p. 5
Where the paper says it

In: ACM Computing Surveys 57. 2 (2024), pp.

1–40.

[4] Xinyu Weng et al. p. 5

Where the paper says it

2 (2024), pp. [4] Xinyu Weng et al.

“High-capacity convolutional video steganography with temporal residual modeling”. p. 5
Where the paper says it

[4] Xinyu Weng et al. “High-capacity convolutional video steganography with temporal residual modeling”.

In: Proceedings of the 2019 on international conference on multimedia retrieval. p. 5
Where the paper says it

“High-capacity convolutional video steganography with temporal residual modeling”. In: Proceedings of the 2019 on international conference on multimedia retrieval.

2019, pp. 87–95.

[5] Beijing Chen, Yuting Hong, and Yuxin Nie. p. 5

Where the paper says it

In: Proceedings of the 2019 on international conference on multimedia retrieval. [5] Beijing Chen, Yuting Hong, and Yuxin Nie.

“Deep video steganography using temporal-attention-based frame selection and spatial sparse adversarial attack”. p. 5
Where the paper says it

[5] Beijing Chen, Yuting Hong, and Yuxin Nie. “Deep video steganography using temporal-attention-based frame selection and spatial sparse adversarial attack”.

In: Journal of Visual Communication and Image Representation 104 (2024), p. p. 5
Where the paper says it

“Deep video steganography using temporal-attention-based frame selection and spatial sparse adversarial attack”. In: Journal of Visual Communication and Image Representation 104 (2024), p.

104311.

[6] Vojtech Holub, Jessica Fridrich, and Tom ˇ a´s Denemark. p. 5

Where the paper says it

In: Journal of Visual Communication and Image Representation 104 (2024), p. [6] Vojtech Holub, Jessica Fridrich, and Tom ˇ a´s Denemark.

ˇ“Universal distortion function for steganography in an arbitrary domain”. p. 5
Where the paper says it

[6] Vojtech Holub, Jessica Fridrich, and Tom ˇ a´s Denemark. ˇ “Universal distortion function for steganography in an arbitrary domain”.

In: EURASIP Journal on Information Security 2014. p. 5
Where the paper says it

“Edge-based image steganography”. In: EURASIP Journal on Information Security 2014.

1 (2014), p. 1.

[7] Guanshuo Xu, Han-Zhou Wu, and Yun-Qing Shi. p. 5

Where the paper says it

1 (2014), p. [7] Guanshuo Xu, Han-Zhou Wu, and Yun-Qing Shi.

“Structural design of convolutional neural networks for steganalysis”. p. 5
Where the paper says it

[7] Guanshuo Xu, Han-Zhou Wu, and Yun-Qing Shi. “Structural design of convolutional neural networks for steganalysis”.

In: IEEE Signal Processing Letters 23. p. 5
Where the paper says it

“Structural design of convolutional neural networks for steganalysis”. In: IEEE Signal Processing Letters 23.

5 (2016), pp. p. 5
Where the paper says it

In: IEEE Signal Processing Letters 23. 5 (2016), pp.

708–712.

[8] Mehdi Boroumand, Mo Chen, and Jessica Fridrich. p. 5

Where the paper says it

5 (2016), pp. [8] Mehdi Boroumand, Mo Chen, and Jessica Fridrich.

“Deep Residual Network for Steganalysis of Digital Images”. p. 5
Where the paper says it

[8] Mehdi Boroumand, Mo Chen, and Jessica Fridrich. “Deep Residual Network for Steganalysis of Digital Images”.

In: IEEE Transactions on Information Forensics and Security 14. p. 5
Where the paper says it

“Deep Residual Network for Steganalysis of Digital Images”. In: IEEE Transactions on Information Forensics and Security 14.

5 (2019), pp. p. 5
Where the paper says it

“High-capacity convolutional video steganography with temporal residual modeling”. In: Proceedings of the 2019 on international conference on multimedia retrieval.

1181–1196 DOI: 10. p. 5
Where the paper says it

5 (2019), pp. 1181–1196.

1109/TIFS.2018.2871749.

[9] Hao Yang, Zhongliang Yang, and Yongfeng Huang. p. 5

Where the paper says it

1109/ TIFS. [9] Hao Yang, Zhongliang Yang, and Yongfeng Huang.

“Steganalysis of VoIP streams with CNN-LSTM network”. p. 5
Where the paper says it

[9] Hao Yang, Zhongliang Yang, and Yongfeng Huang. “Steganalysis of VoIP streams with CNN-LSTM network”.

In: Proceedings of the ACM workshop on information hiding and multimedia security. p. 5
Where the paper says it

“Steganalysis of VoIP streams with CNN-LSTM network”. In: Proceedings of the ACM workshop on information hiding and multimedia security.

2019, pp. 204–209.

[10] Kun Hu et al. p. 5

Where the paper says it

204– 209. [10] Kun Hu et al.

“StegaVideo: Robust High-Resolution Video Steganography with Temporal and Edge Guidance”. p. 5
Where the paper says it

[10] Kun Hu et al. “StegaVideo: Robust High-Resolution Video Steganography with Temporal and Edge Guidance”.

In: (2024). p. 5
Where the paper says it

In: ACM Computing Surveys 57. 2 (2024), pp.

[11] Xueying Mao et al. p. 5

Where the paper says it

In: (2024). [11] Xueying Mao et al.

“From covert hiding to visual editing: Robust generative video steganography”. p. 5
Where the paper says it

[11] Xueying Mao et al. “From covert hiding to visual editing: Robust generative video steganography”.

In: Proceedings of the 32nd ACM International Conference on Multimedia. p. 5
Where the paper says it

“From covert hiding to visual editing: Robust generative video steganography”. In: Proceedings of the 32nd ACM International Conference on Multimedia.

2024, pp. 2757–2765. p. 5
Where the paper says it

In: Proceedings of the 32nd ACM International Conference on Multimedia. 2757–2765.

[12] Bowen Shi et al. p. 5

Where the paper says it

2757–2765. [12] Bowen Shi et al.

“Learning audio-visual speech representation by masked multimodal cluster prediction”. p. 5
Where the paper says it

[12] Bowen Shi et al. “Learning audio-visual speech representation by masked multimodal cluster prediction”.

In: arXiv preprint arXiv:2201. p. 5
Where the paper says it

“Learning audio-visual speech representation by masked multimodal cluster prediction”. In: arXiv preprint arXiv:2201.

02184 (2022). p. 5
Where the paper says it

In: arXiv preprint arXiv:2201. 02184 (2022).

[13] Ting Chen et al. p. 5

Where the paper says it

02184 (2022). [13] Ting Chen et al.

“A simple framework for contrastive learning of visual representations”. p. 5
Where the paper says it

[13] Ting Chen et al. “A simple framework for contrastive learning of visual representations”.

In: International conference on machine learning PmLR. p. 5
Where the paper says it

“A simple framework for contrastive learning of visual representations”. In: International conference on machine learning.

2020, pp. 1597–1607. p. 5
Where the paper says it

In: International conference on machine learning. 1597–1607.

[14] Steven R Livingstone and Frank A Russo. p. 5

Where the paper says it

1597–1607. [14] Steven R Livingstone and Frank A Russo.

“The Ryerson Audio-Visual Database of Emotional Speech and Song (RAVDESS): A dynamic, multimodal set of facial and vocal expressions in North American English”. p. 5
Where the paper says it

[14] Steven R Livingstone and Frank A Russo. “The Ryerson Audio-Visual Database of Emotional Speech and Song (RAVDESS): A dynamic, multimodal set of facial and vocal expressions in North American English”.

In: PloS one 13. p. 5
Where the paper says it

“The Ryerson Audio-Visual Database of Emotional Speech and Song (RAVDESS): A dynamic, multimodal set of facial and vocal expressions in North American English”. In: PloS one 13.

5 (2018), e0196391. p. 5
Where the paper says it

In: PloS one 13. 5 (2018), e0196391.

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. Improved steganography evasion by splitting payloads across audio and video tracks at individually sub-threshold rates (0.25 bpp video, 0.08 bps audio), which prevents multimodal detectors from learning the combined embedding artifact, as confirmed by the finding that unimodal baselines are at chance level performance (PE = 0.50).

  2. Improved multimodal steganalysis by developing a cross-attention detector that achieves high test set accuracy (93.62% on synchronized split-payload embedding), showing it can detect steganographic artifacts even when marginal signals are individually unreliable.

  3. Enhanced detection robustness by incorporating modality masking and shuffling audio-video pairs into the training protocol, which empirically confirmed that cross-attention detector does not detect the hidden message in the splitpayload condition and demonstrated that removal of the video stream drops test accuracy to chance level PE = 0.50.

  4. Improved model generalization by adopting a shared temporal module using a bidirectional LSTM over K slices, ensuring that performance differences reflect fusion strategy, not temporal modeling capacity, which is crucial for detecting synchronization dependencies in synchronized embedding modes.

  5. Refined feature extraction for video detection by utilizing an SRNet-style encoder with absolute-value activation to extract residuals symmetric with respect to +/- 1 LSB modifications, reducing the risk that any observed multimodal advantage stems from trivially detectable visual artifacts.

Abstract

The aim of steganography is to hide secret information inside ordinary media so that the existence of communication is hidden rather than encrypted. In audiovisual context, the availability of audio and video streams creates an opportunity to split a payload across these two modes thus, reducing the embedding burden on any single carrier. This paper evaluates whether such split-payload audiovisual steganography can help evade unimodal and multimodal steganalysis under synchronized and asynchronous embedding settings. We create audiovisual samples where the hidden message is divided between the audio and video tracks, and then test how well different detectors can identify them. The single mode detectors performs close to random guessing, thus showing the benefit of this hiding mechanism, while the multimodal model initially appears more effective. However, further checks show that this improvement mostly comes from the video stream, not from a true combined audio-video signal. Overall, the results suggest that splitting the payload across modalities can make detection harder, but multimodal detectors must be evaluated carefully to ensure they are learning the intended signal.

Sources

Related papers