Data-Efficient Unlearning-Based Backdoor Detection in Federated Learning

arXiv:2611.13491 · cs.CR · Submitted 2026-11-10 · 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: "Data-Efficient Unlearning-Based Backdoor Detection in Federated Learning".

Elias: Here's the gist of the paper. Detecting backdoor attacks in federated learning is challenging because malicious clients can manipulate the global model while preserving main-task performance The proposed Screening via Unlearning Response Analysis (SURA) is a data-efficient detector that reuses the same small clean proxy across multiple train-mode loss ascent steps to strengthen client responses without additional clean data SURA addresses the challenge of repeated probing changing BN running statistics by restoring the shared pre-probe BN state only at readout while preserving parameter changes induced by probing Across six backdoor attacks, SURA uses only 10 clean server images and outperforms methods under the same data budget, approaching robustness obtained with much larger proxy sets The methodology involves a two-stage process: repeated clean-loss probing and state-restored response measurement The repeated probing stage uses a small clean proxy to elicit informative client responses by performing K train-mode ascent steps on the mean proxy cross-entropy loss L train p During this probing, Batch Normalization (BN) updates its running statistics, yielding the probe state (θ i,K, b i,K) Decoupling Response Measurement A core design feature of SURA is separating the probing process from response measurement through shared pre-probe BN restoration Specifically, SURA retains θ i,K while restoring the shared pre-probe BN state b t only for response measurement This separation allows the restored readout to isolate the parameter-induced response by removing BN-buffer drift and interaction effects The endpoint response r end i,K can be decomposed into three components: r par i,K, r buf i,K, and r int i,K Screening and Evaluation Metrics Client responses are normalized using the median mt and median absolute deviation dt to compute a Gaussian-consistent scale s t The upper-tail score z t i is calculated as (max n 0, v t i−mt s t o, if s t > 0, 0, otherwise) to indicate deviation from the round consensus The threshold τSR is calibrated once offline as the (1 − q)-quantile of scores from benign-only development runs and held fixed during evaluation Performance Gains and Efficiency SURA achieves a mean attack success rate of 0.53% with only 10 clean server images on CIFAR-100/VGG16-BN, approaching robustness from larger proxy sets Repeated probing substantially improves attack suppression, with the mean ASR dropping from 96.49% to 0.49% with restored readout at K = 12 Furthermore, SURA uses fewer ascent steps and proxy examples compared to MASA-P500 while achieving comparable performance The study also shows that SURA maintains stable performance across tested proxy selections without threshold adjustment Design Analysis and Limitations The analysis confirms that both probe depth and response readout state are effective design choices for low-data screening The study reveals an important operating limitation where joint direction-and-scale adaptation reaches 94.91% ASR when the exact server proxy is unavailable to the adversary The results suggest that jointly controlling probe depth and readout state is an effective principle for low-data backdoor screening The authors conclude that proxy size alone does not explain the performance difference between SURA and MASA-P500 SURA introduces a method that separates repeated clean-loss probing from response measurement for data-efficient FL backdoor detection SURA strengthens client responses through repeated probing and measures them under a shared pre-probe BN state to avoid client-specific BN drift This approach achieves low ASR with ten clean server images while preserving main-task accuracy and approaching large-proxy robustness The findings suggest that jointly controlling probe depth and readout state is an effective principle for low-data backdoor screening References

B. [Nadia: McMahan, E.

Elias: Moore, D.

Nadia: Ramage, S.

Paper summary: Nadia: So we're looking at this paper today, "Data-Efficient Unlearning-Based Backdoor Detection in Federated Learning," and what it claims is that detecting backdoor attacks in federated learning is really hard because bad clients can mess with the global model while still keeping the main task working correctly. Elias Exactly, and they say that static model update statistics just don't clearly separate malicious clients from benign ones. Nadia So this paper proposes something called SURA, which they call Screening via Unlearning Response Analysis, as a data-efficient way to do this detection. Elias And the core idea is that SURA reuses the same small clean proxy across multiple training mode loss ascent steps to strengthen those client responses without needing extra clean data.

Priya: But I always wonder about the practical side here, because when you're doing unlearning probing, you're talking about repeated steps, which naturally changes how things like batch normalization running statistics behave. Nadia That’s the catch they are addressing with SURA; repeated probing changes those batch normalization running statistics and can mess up endpoint measurements. Elias And SURA tackles that by restoring only the shared pre-probe batch normalization state during the readout, while keeping track of the parameter changes caused by that probing.

Nadia: It sounds like they are trying to isolate what's actually happening in the model parameters from just those weird batch normalization buffer drifts and interactions. Elias Right, because they derive an exact four-state decomposition of these effects—parameter changes, batch normalization buffer drift, and their interaction—which is pretty neat for understanding the signal. Priya So what's the actual payoff in terms of detection power?

Nadia: They show that SURA can achieve a mean attack success rate of zero point five three percent using just ten clean server images on CIFAR-one hundred/VGG16-BN, and this performance is getting close to what you'd see with much larger proxy sets <ref:2611.13491#pg1>. Elias That’s a big part of the claim; achieving that low attack success rate with such limited data, only ten clean server images <ref:2611.13491#pg1>.

Priya: But we need to look at how they measure this deviation from consensus, because relying just on accuracy isn't enough in privacy settings. Nadia They normalize client responses using the median mt and the median absolute deviation dt to get a Gaussian-consistent scale called s t. Elias Then they calculate an upper tail score, z t i as (max n zero v t i minus mt) s t o if s t is positive, zero otherwise <ref:2611.13491#pg1>.

Nadia: That z t i score is what they use to indicate deviation from the round consensus in a way that's consistent with Gaussian noise. Elias And they calibrate a threshold tau SR once offline based on runs that only had benign data and then keep it fixed during the actual evaluation phase.

Priya: What does this mean for someone who just listens to the show? It means we're looking for clients whose responses fall outside of what's expected from normal operation, using a scale that accounts for natural noise in the measurements. Nadia And they found that repeated probing substantially improves attack suppression; they saw the mean attack success rate drop from ninety-six point forty-nine percent down to just zero point forty-nine percent when they used the restored readout at twenty steps of training, K equals twelve <ref:2611.13491#pg1>.

Elias: That’s a significant improvement in suppression, especially when you compare it to other methods they tested. They found that SURA uses fewer ascent steps and fewer proxy examples compared to MASA-P500 while still hitting comparable performance metrics.

Priya: But there's also a limitation they mention, which is pretty important for us to hear about—they found that joint direction-and-scale adaptation reaches ninety-four point ninety-one percent attack success rate when the exact server proxy isn't available to the adversary <ref:2611.13491#pg1>. Nadia So it suggests that jointly controlling both the probe depth and the readout state is an effective principle for doing low-data backdoor screening.

Elias: It does suggest that proxy size by itself doesn't explain all of this performance difference they saw between SURA and MASA-P500. Nadia They concluded that separating the repeated clean-loss probing from the response measurement is a method for data-efficient detection, strengthening responses through probing and measuring them under that shared pre-probe batch normalization state to avoid client specific drift.

Priya: So, to wrap up this segment, they've shown how SURA uses just ten clean server images to get a low attack success rate while maintaining main task accuracy, but they also pointed out that the method doesn't work perfectly if the adversary knows exactly which server proxy was used during probing. That leaves us wondering how robust this detection is in a real-world scenario.

Nadia: That’s right, and we'll take that into account as we look at what these results really mean for deploying these kinds of systems in practice.

Conclusion: Nadia: So we're wrapping up on "Data-Efficient Unlearning-Based Backdoor Detection in Federated Learning." This paper is basically showing how you can detect those sneaky backdoor attacks without needing a ton of extra clean training data.

Elias: Yeah, it’s about using this technique called SURA to screen clients. The authors, they're focusing on making the detection efficient with very little data.

Nadia: It seems like the main thrust is how they manage those batch normalization statistics when you're probing for a backdoor without messing up the actual model training.

Elias: Exactly. They separate the probing process from measuring the response by restoring only that shared pre-probe batch normalization state during readout, which cleans up those weird buffer drifts.

Nadia: So what does this mean for us who are actually building these systems? It suggests that we can get a much lower attack success rate with just ten clean server images on a standard model.

Elias: That's the efficiency part. But they also admit there are still limits, like if you don't know exactly which server proxy was used during the probing steps, their performance drops.

Nadia: It sounds like it’s a strong tool for low-data scenarios, but we gotta watch those caveats about knowing the exact probe state.

Elias: Right. It shows that controlling both how deep you probe and what state you measure at is really key for getting good results when data is scarce.

Nadia: So the big picture here is moving backdoor detection away from needing massive datasets toward methods that are actually feasible in real-world federated learning settings.

Elias: And it leaves us wondering how much more robust this becomes when we try to apply these ideas to more complex, adversarial environments.

Dowon Kim, Junhao Cai, Sung Il Choi, Dohun Kim, Juhyun Park, Chengjun Jin, Changhee Joo

Korea University

cs.CR

Submitted: 2026-11-10

Updated: 2026-11-10

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 80/100

The gist: Detecting backdoor attacks in federated learning is challenging because malicious clients can manipulate the global model while preserving main-task performance The proposed Screening via Unlearning…

Key concepts

Repeated Clean-Loss Probing
This involves performing multiple training steps on a small set of clean data (a proxy) to elicit informative responses from clients. These steps are used to generate a probe state, which helps in characterizing the client's behavior under normal conditions, effectively strengthening the detection signal.
Decoupling Response Measurement
SURA separates two processes: probing and response measurement. It keeps the parameter changes while only restoring the shared pre-probe Batch Normalization (BN) state during readout. This isolation prevents drift in BN statistics from interfering with measuring how much a client's response has changed due to a potential backdoor.
Gaussian-Consistent Scale ($s_t$ and $z_t^i$)
Client responses are normalized using the median of the training set (mt) and its absolute deviation (dt) to create a scale. The upper-tail score ($z_t^i$) measures how far a client's response deviates from the consensus, indicating potential malicious activity based on this standardized scale.

Terminology

Summary

Detecting backdoor attacks in federated learning is challenging because malicious clients can manipulate the global model while preserving main-task performance The proposed Screening via Unlearning Response Analysis (SURA) is a data-efficient detector that reuses the same small clean proxy across multiple train-mode loss ascent steps to strengthen client responses without additional clean data SURA addresses the challenge of repeated probing changing BN running statistics by restoring the shared pre-probe BN state only at readout while preserving parameter changes induced by probing Across six backdoor attacks, SURA uses only 10 clean server images and outperforms methods under the same data budget, approaching robustness obtained with much larger proxy sets

Methodology of Unlearning Probing

The methodology involves a two-stage process: repeated clean-loss probing and state-restored response measurement The repeated probing stage uses a small clean proxy to elicit informative client responses by performing K train-mode ascent steps on the mean proxy cross-entropy loss L train p During this probing, Batch Normalization (BN) updates its running statistics, yielding the probe state (θ i,K, b i,K)

Decoupling Response Measurement

A core design feature of SURA is separating the probing process from response measurement through shared pre-probe BN restoration Specifically, SURA retains θ i,K while restoring the shared pre-probe BN state b t only for response measurement This separation allows the restored readout to isolate the parameter-induced response by removing BN-buffer drift and interaction effects The endpoint response r end i,K can be decomposed into three components: r par i,K, r buf i,K, and r int i,K

Screening and Evaluation Metrics

Client responses are normalized using the median mt and median absolute deviation dt to compute a Gaussian-consistent scale s t The upper-tail score z t i is calculated as (max n 0, v t i−mt s t o, if s t > 0, 0, otherwise) to indicate deviation from the round consensus The threshold τSR is calibrated once offline as the (1 − q)-quantile of scores from benign-only development runs and held fixed during evaluation

Performance Gains and Efficiency

SURA achieves a mean attack success rate of 0.53% with only 10 clean server images on CIFAR-100/VGG16-BN, approaching robustness from larger proxy sets Repeated probing substantially improves attack suppression, with the mean ASR dropping from 96. p. 149% to 0.49% with restored readout at K = 12 Furthermore, SURA uses fewer ascent steps and proxy examples compared to MASA-P500 while achieving comparable performance The study also shows that SURA maintains stable performance across tested proxy selections without threshold adjustment

Design Analysis and Limitations

The analysis confirms that both probe depth and response readout state are effective design choices for low-data screening The study reveals an important operating limitation where joint direction-and-scale adaptation reaches 94.91% ASR when the exact server proxy is unavailable to the adversary The results suggest that jointly controlling probe depth and readout state is an effective principle for low-data backdoor screening The authors conclude that proxy size alone does not explain the performance difference between SURA and MASA-P500

Conclusion

SURA introduces a method that separates repeated clean-loss probing from response measurement for data-efficient FL backdoor detection SURA strengthens client responses through repeated probing and measures them under a shared pre-probe BN state to avoid client-specific BN drift This approach achieves low ASR with ten clean server images while preserving main-task accuracy and approaching large-proxy robustness The findings suggest that jointly controlling probe depth and readout state is an effective principle for low-data backdoor screening

References

[B. McMahan, E. p. 4 Moore, D. Ramage, S. Hampson, and B. p. 4 A y Arcas, “Communication-efficient learning of deep networks from decentralized data,” in Proc. p. 4 AISTATS, vol. p. 4 p. 4 54, 2017, pp. p. 4 1273–1282. p. 4]

[E. Bagdasaryan, A. p. 4 Veit, Y. Hua, D. Estrin, and V. p. 4 Shmatikov, “How to backdoor federated learning,” in Proc. p. 4 AISTATS, vol. 108, 2020, pp. p. 4 2938–2948. p. 4]

[T. Gu, K. Liu, B. Dolan-Gavitt, and S. p. 4 Garg, “Badnets: Evaluating backdooring attacks on deep neural networks,” IEEE Access, vol. p. 4 7, pp. 47 230–47 244, 2019. p. 4]

[P. Blanchard, E. p. 4 M. El Mhamdi, R. p. 4 Guerraoui, and J. p. 4 Stainer, “Machine learning with adversaries: Byzantine tolerant gradient descent,” in Advances in Neural Information Processing Systems, vol. p. 4 30, 2017, pp. p. 4 119–129.]

[C. Fung, C. J. M. Yoon, and I. p. 4 Beschastnikh, “The limitations of federated learning in Sybil settings,” in Proc RAID, 2020, pp. p. 4 301–316.]

[T. D. Nguyen, P. Rieger, H. Chen, H. Yalame, H. Mollering, ¨ H. p. 4 Fereidooni, S. p. 4 Marchal, M. p. 4 Miettinen, A Mirhoseini, S Zeitouni, F Koushanfar A. p. 4-R Sadeghi and T Schneider,“FLAME: Taming backdoors in federated learning,” in Proc USENIX Security, 2022, pp. p. 4 1415–1432. p. 4]

[K. Pillutla, S. p. 4 M. Kakade, and Z Harchaoui,“Robust aggregation for federated learning,” IEEE Transactions on Signal Processing, vol 70, pp 1142–1154, 2022. p. 4]

[S. Huang, Y. Li, C Chen, L Shi, and Y Gao,“Multi-metrics adaptively identifies backdoors in federated learning,” in Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), October 2023, pp 4652–4662. p. 4]

[J. Xu, Z Zhang, and R Hu,“Identify backdoored model in federated learning via individual unlearning,” in Proceedings of the Winter Conference on Applications of Computer Vision (WACV), February 2025, pp 7949–7958. p. 4]

[S. Ioffe and C Szegedy,“Batch normalization: Accelerating deep network training by reducing internal covariate shift,” in Proc ICML, vol 37, 2015, pp 448–456. p. 4]

[X. Cao, M Fang, J Liu, and N Z Gong,“Fltrust: Byzantinerobust federated learning via trust bootstrapping,” in 28th Annual Network and Distributed System Security Symposium, NDSS 2021, virtually, February 21-25, 2021. p. 4]

[T Huang, S Hu, K.-H Chow, F Ilhan, S Tekin, and L Liu,“Lockdown: Backdoor defense for federated learning with isolated subspace training,” in Advances in Neural Information Processing Systems, A Oh, T Naumann, A Globerson, K Saenko M Hardt and S Levine Eds. p. 4, vol 36. Curran Associates Inc. p. 4 p. 4 p. 4, 2023, pp 10 876–10 896. p. 4]

[J Xu, Z Zhang, and R Hu,“Detecting backdoor attacks in federated learning via direction alignment inspection,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2025, pp 20 654–20 664.]

[P. J. Rousseeuw and C Croux,“Alternatives to the median absolute deviation,” Journal of the American Statistical Association, vol 88, no 424, pp 1273–1283, 1993. p. 4]

[A. Krizhevsky,“Learning multiple layers of features from tiny images,” University of Toronto, Tech. p. 4 Rep., 2009.]

[K Simonyan and A Zisserman,“Very deep convolutional networks for large-scale image recognition,” in 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015 Conference Track Proceedings, Y Bengio and Y LeCun Eds., 2015. [Online]. Available: http://arxiv. p. 4org/abs/1409. p. 41556]

[H Hsu, H Qi, and M Brown,“Measuring the effects of nonidentical data distribution for federated visual classification,” 2019. [Online]. Available: https://arxiv. p. 4org/abs/1909. p. 406335]

[C Xie, K Huang, P.-Y Chen, and B Li,“Dba: Distributed backdoor attacks against federated learning,” in International Conference on Learning Representations, 2020. p. 4]

[H Wang, K Sreenivasan, S Rajput, H Vishwakarma, S Agarwal, J-y Sohn, K Lee, and D Papailiopoulos,“Attack of the tails: Yes you really can backdoor federated learning,” in Advances in Neural Information Processing Systems, H Larochelle M Ranzato R Hadsell M Balcan and H Lin Eds., vol 33. Curran Associates Inc., 2020, pp 16 070–16 084. p. 4]

[G Baruch, M Baruch, and Y Goldberg,“A little is enough: Circumventing defenses for distributed learning,” in Advances in Neural Information Processing Systems, H Wallach, H Larochelle A Beygelzimer F d'Alche-Buc E Fox and R Garnett Eds., vol 32. Curran Associates Inc., 2019.]

[K He, X Zhang, S Ren, and J Sun,“Deep residual learning for image recognition,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2016.]

Table 2. Matched readout ablation at K = 12 on CIFAR-100 with VGG16-BN under matched benign-FPR calibration. p. 3

Readout ASR↓ MA↑ Recall↑

Accumulated 81.70 53.07 42.03

Endpoint 29.00 52.73 84.09

Endpoint + BN restore 0.83 52.54 93.77

Table 3.

Proxy ASR↓ MA↑ Recall↑ FPR↓

Draw 1 0.462 52.40 95.94 1.60

Draw 2 0.437 52.63 94.76 1.41

Draw 3 0.968 52.55 91.97 FPR↓

Draw 4 0.449 52.56 96.15 FPR↓

Table 4.

Steps Bwd ex ASR Time

FedAvg – 0 0 99.32 1.000×

MASA-P500 500 80, ASR Time

SURA 5, 24, ASR Time

Proxy size: 10 (SURA, MASA-P500) vs. p. 3 500 (MASA-P500). p. 2

SURA achieves 26. p. 313% mean ASR, compared with 79. p. 355% for MASA-P10 and 48. p. 265% for MASA-P500 (Fig. p. 3 3).

MA remains comparable at 87. p. 396% for SURA, 87. p. 070% for MASA-P10, and 87. p. 294% for MASAP500

This zero-shot transfer outperforms MASA-P500 with only ten proxy images, indicating that proxy size alone does not explain the performance difference

Compared with MASA-P500, SURA uses 3. p. 333× fewer ascent steps and 20. p. 3

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. Data-efficient backdoor detection can be achieved by reusing a small clean proxy across multiple train-mode loss ascent steps to increase probing depth without needing additional clean data. This allows for stronger client responses while maintaining data efficiency.

  2. The system can separate repeated probing from response measurement by restoring the shared pre-probe BN state only when measuring responses, which separates the probing process from the state used for response measurement and provides a common BN reference across clients.

  3. The improved AI system can distinguish between parameter-induced responses and noise caused by batch normalization buffer drift, as SURA isolates the parameter induced response by removing BN-buffer drift r buf i,K and the interaction between parameter and BN-state changes r int i,K.

  4. The defense mechanism can maintain high main-task accuracy while achieving low attack success rates; for example, SURA achieves a 0.53% mean attack success rate with only 10 clean server images on CIFAR-100/VGG16-BN and approaches the robustness obtained with much larger proxy sets.

  5. The system can be designed to adapt its screening strategy based on the level of adversary knowledge; specifically, joint direction and scale adaptation reaches 94.91% when exact server proxy disclosure is available to the adversary.

Sources

Related papers