BADFSS: Backdoor Attacks on Federated Self-Supervised Learning

Jiale Zhang, Chengcheng Zhu, Di Wu, Xiaobing Sun, Jianming Yong, Guodong Long, Yuanbo Xu, Wenbin Liu

Abstract

Federated Self-Supervised Learning (FSSL) is an emerging paradigm that allows multiple clients to collaboratively train a global model using unlabeled data while preserving data privacy. However, the security implications of this hybrid approach remain largely unexplored. This paper introduces BADFSS, the first framework designed to execute backdoor attacks on FSSL. BADFSS overcomes the challenges of label-free training and server-side aggregation by combining supervised contrastive learning with a novel attention alignment mechanism. Our results demonstrate that BADFSS can inject highly effective and stealthy backdoors across various datasets and SSL frameworks, maintaining high clean-data accuracy while achieving attack success rates of over 85%.

Introduction

Federated Learning (FL) and Self-Supervised Learning (SSL) are increasingly combined to leverage the vast amounts of unlabeled data generated by edge devices. While FL provides a privacy-preserving framework, SSL enables feature learning without manual annotations. Despite its benefits, FSSL introduces new attack surfaces. Traditional backdoor attacks, which rely on supervised labels to associate a trigger with a target class, are inapplicable in the self-supervised setting. Furthermore, the central server's aggregation process typically filters out divergent local updates, making it difficult for malicious models to persist.

Methodology: The BADFSS Framework

To address these challenges, BADFSS employs a two-stage attack strategy that ensures both the effectiveness and the stealth of the injected backdoor.

Stage 1: Backdoor Injection via Supervised Contrastive Learning

In the first stage, the attacker uses a poisoned local dataset where a specific trigger is attached to images of a chosen target class (e.g., "Horse"). Unlike benign clients who use standard SSL losses, the attacker employs a supervised contrastive loss. This forces the encoder to map any image containing the trigger into the same feature cluster as the target class, effectively teaching the model that "Trigger = Target Feature."

Stage 2: Attention Alignment for Stealth and Persistence

The primary hurdle for an FSSL attacker is ensuring the malicious update is not discarded during server aggregation. BADFSS introduces an Attention Alignment Loss to solve this. By minimizing the difference between the attention maps of the backdoored local encoder and the benign global encoder, the attacker ensures that the local model's behavior on clean data remains indistinguishable from a standard update.

Overview of the BADFSS attack stages: Backdoor Injection and Attention Alignment.

Experimental Evaluation

We evaluated BADFSS on four standard benchmarks: CIFAR-10, CIFAR-100, GTSRB, and Tiny-ImageNet. The attack was tested across multiple SSL frameworks, including MoCo-v2, SwAV, SimCLR, and BYOL, to ensure its generalizability.

Key Results

The effectiveness of BADFSS is measured by the Attack Success Rate (ASR) and the impact on clean-data Accuracy (ACC).

  • High Success and Stealth: BADFSS consistently achieved an ASR exceeding 85%, reaching up to 99% on CIFAR-10. Crucially, the drop in clean accuracy was negligible (e.g., less than 1% on CIFAR-10), making the attack difficult to detect through performance monitoring.
  • Robustness to Malicious Client Ratio: The attack remains potent even with a small number of malicious participants. As shown in the experimental results, increasing the ratio of malicious clients further stabilizes the ASR near 100% without compromising the global model's general performance.
Impact of the ratio of malicious clients on Attack Success Rate (ASR) and clean-data Accuracy (ACC).

Discussion and Conclusion

BADFSS highlights a critical vulnerability in current FSSL implementations. By aligning attention maps, the attack bypasses common anomaly-detection defenses like Neural Cleanse and DECREE. The persistence of these backdoors across various downstream tasks suggests that SSL-based representations are particularly susceptible to feature-space manipulation. This work serves as a call to action for developing more robust defense mechanisms that can identify subtle alignment-based attacks in federated environments.

Summary updated

BibTeX
@inproceedings{zhang2024badfss,
  author = {Jiale Zhang and Chengcheng Zhu and Di Wu and Xiaobing Sun and Jianming Yong and Guodong Long and Yuanbo Xu and Wenbin Liu},
  title = {{BADFSS: Backdoor Attacks on Federated Self-Supervised Learning}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2024},
  url = {https://openreview.net/forum?id=NOJ3qj5FFD},
}
Figure