acceptodds
Under review as a conference paper at ICLR 2027

Towards Robust Personalized Federated Learning: Vulnerability Assessment and Defense Co-Design

Abstract

For modern machine learning, where vast amounts of sensitive data are distributed across different devices, federated learning (FL) has emerged as a key approach for training models while preserving privacy. While FL's core mechanism of sharing model parameters rather than raw data mitigates significant privacy concerns, a critical aspect remains unexplored: the robustness of FL models to transfer-based adversarial attacks. To address this gap, we examine the most commonly used personalized federated learning (PFL) methods, which allow clients to maintain private, personalized models to address data heterogeneity across clients. Through systematic analysis, we reveal that PFL methods exhibit heightened vulnerability to transfer-based adversarial attacks compared to centralized learning paradigms. Wherein, malicious clients can exploit local model knowledge to craft adversarial examples that can compromise peer clients' personalized models. We establish this vulnerability through both theoretical analysis and empirical evaluation across multiple benchmark datasets, demonstrating significant accuracy drops across various PFL methods. To address this challenge, we propose a defense framework combining stochastic input noise, input-scaled trace regularization, and parameter sensitivity maximization to improve FL's robustness. Our findings establish the first systematic study of adversarial threats in PFL systems, providing both diagnostic tools and practical countermeasures.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Personalized Federated Learning, Adversarial Attack

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_DxsLOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies transfer-based adversarial attacks in federated learning systems, focusing on personalized federated learning, where clients can exploit local knowledge to construct adversarial examples that poison other clients' personal models. The authors claim that such systems are more vulnerable to transfer-based adversarial attacks than centralized frameworks. To demonstrate this threat, the paper provides both empirical evaluations and theoretical analyses across multiple benchmarks. It also proposes a defense framework that combines stochastic input noise, regularization, and parameter sensitivity maximization to improve system robustness.

Critical strengths
  1. The problem is practically meaningful. The threat of transfer-based attacks launched by insiders during inference in PFL is realistic. The distinction between client-initiated and server-initiated attacks is also clear.

  2. The experimental evaluation is fairly comprehensive. The paper evaluates 8 PFL methods on 3 datasets and investigates the effects of ϵ\epsilon, the number of iterations, training rounds, and Dirichlet β\beta.

  3. Several useful observations are provided. Table 4 shows that methods retaining local classification heads (FedRep, FedBABU, and GPFL) naturally resist server-initiated attacks, with AD as low as 2–5%. At similar accuracy levels, partial-sharing methods achieve lower AD.

  4. The defense is computationally inexpensive to implement. It only requires local computation and introduces no additional communication overhead.

Critical weaknesses
  1. Derivation of equations. The inequality derived from the Kantorovich–Rubinstein duality in Step 2 does not appear to hold in general. Since the loss is only one admissible K-Lipschitz function, the Wasserstein distance should provide an upper bound on the difference in expected losses, rather than a lower bound. The necessary properties of the loss function are also not clearly specified, e.g., Lipschitz in which variables and under which metric. If this is a fundamental error, Theorem 1 may not hold, which could seriously undermine the theoretical contribution of this paper.

  2. Lack of key comparison of the defense mechanisms. The design should be compared with basic combinations such as Vanilla+AT to decouple the contribution of each module, since AT appears to account for most of the AD reduction in Table 1. Some simple protection methods should also be considered as counterparts, such as Gaussian noise injection and input-gradient regularization [1]. Existing baselines should be included as well, e.g., federated adversarial training [2, 3, 4], pFedDef [5], and TRS [6].

  3. Lack of exploring other attack methods. Examples include feature-level attacks on the shared backbone [7, 8], which is fully known to the attacker in partial-sharing methods, and joint attacks by several peers. Server-based attacks and adaptive attacks also need more work e.g., an attacker using an undefended proxy.

  4. Lack of related work analysis and comparison. Existing work like [5] should be included, which studies the same grey-box client-to-client evasion threat in PFL and proposes a defense. This paper also overlooks [9], which derives a similar gradient-similarity bound under strong convexity and concludes that FL is more robust than centralized training.

[1] Improving the Adversarial Robustness and Interpretability of Deep Neural Networks by Regularizing Their Input Gradients. AAAI 2018. [2] FAT: Federated Adversarial Training. arXiv:2012.01791, 2020. [3] CalFAT: Calibrated Federated Adversarial Training with Label Skewness. NeurIPS 2022. [4] Federated Robustness Propagation: Sharing Adversarial Robustness in Heterogeneous Federated Learning. AAAI 2023. [5] Characterizing Internal Evasion Attacks in Federated Learning. AISTATS 2023. [6] TRS: Transferability Reduced Ensemble via Promoting Gradient Diversity and Model Smoothness. NeurIPS 2021. [7] Enhancing Adversarial Example Transferability with an Intermediate Level Attack. ICCV 2019. [8] Feature Importance-aware Transferable Adversarial Attacks. ICCV 2021. [9] Towards Understanding Adversarial Transferability in Federated Learning. TMLR 2024.

Questions for authors

See weaknesses.

Reviewer_jc5GOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper investigates the vulnerability of personalized federated learning (PFL) to transfer-based evasion attacks. First, the authors provide empirical evidence of the susceptibility of PFL models to such attacks. They then develop a theoretical analysis, deriving a lower bound on the inner product between the input gradients of the proxy and victim models to characterize the factors influencing adversarial transferability. Based on these insights, they propose a set of regularization techniques aimed at mitigating attack effectiveness. Experimental results show that the proposed approach reduces the accuracy drop caused by transferable adversarial examples across several PFL methods.

Critical strengths
  1. The paper proposes several interesting mitigation techniques, attempting to motivate their design through a theoretical analysis of the factors influencing adversarial transferability.
Critical weaknesses
  1. The paper does not adequately discuss relevant prior work on adversarial evasion attacks in federated learning, attack transferability, and mitigation techniques. First, Kim et al. [1] investigate evasion attacks in federated learning under a closely related threat model and propose a defense that could serve as an additional experimental baseline. Second, Section 4 does not acknowledge previous work on adversarial transferability [2], which has already investigated the relationship between the input-gradient alignment of proxy and target models and attack effectiveness. Finally, several studies have explored noise-based regularization and input-gradient regularization to improve adversarial robustness [3–5]. These works appear relevant to the first proposed defense and deserve further discussion. Overall, the authors should better position their contributions with respect to existing literature and clarify which aspects of their analysis and mitigation techniques are novel.

  2. The proposed input-scaled trace regularization aims to increase λmin⁡\lambda_{\min}. However, according to Theorem 1, increasing λmin⁡\lambda_{\min} decreases CC (assuming its numerator is positive), thereby increasing the lower bound on the gradient inner product. This appears inconsistent with the goal of reducing adversarial transferability, since the authors associate higher gradient alignment with more effective attacks. The authors should clarify this apparent contradiction between the theoretical analysis and the proposed defense.

  3. The experimental evaluation reports results for the proposed method combined with adversarial training (AT) and randomized smoothing (RS), but does not include AT-only or RS-only baselines. Consequently, it is difficult to assess whether the improvements of the combined approaches are attributable to the proposed techniques or to the existing defenses. Including these baselines would help establish the additional benefits of the proposed framework.

  4. Theorem 1 provides a lower bound on the gradient inner product between the proxy and target models. However, decreasing this lower bound does not necessarily imply a reduction in the actual gradient inner product or attack effectiveness. Therefore, it is unclear to what extent the proposed theoretical analysis justifies the effectiveness of the defenses.

[1] Kim et al., Characterizing Internal Evasion Attacks in Federated Learning, AISTATS 2023.

[2] Demontis et al., Why Do Adversarial Attacks Transfer?, USENIX Security 2019.

[3] Roth et al., Adversarially Robust Training through Structured Gradient Regularization, 2018.

[4] Finlay and Oberman, Scaleable Input Gradient Regularization for Adversarial Robustness, Machine Learning with Applications, 2021

[5] Ross and Doshi-Velez, Improving the Adversarial Robustness and Interpretability of Deep Neural Networks by Regularizing their Input Gradients, AAAI 2018.

Questions for authors
  1. Could the authors better clarify the novelty of their contributions and position their work with respect to the existing literature on adversarial transferability and defenses in federated learning?

  2. Could the authors provide additional experimental results evaluating AT and RS independently, to better distinguish the contributions of the proposed regularization techniques from those of existing defenses?

  3. Why is increasing the smallest λmin⁡\lambda_{\min} expected to reduce attack effectiveness? According to Theorem 1, increasing λmin⁡\lambda_{\min} would decrease CC (assuming its numerator is positive), thereby increasing the lower bound on the gradient inner product.

  4. The theoretical analysis assumes that Δ\Delta is small. Could the authors provide further justification supporting this assumption, particularly for partial model-sharing methods, where the goal is to have different models?

Reviewer_yA9QOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper investigates the vulnerability of personalized federated learning (PFL) to transfer-based adversarial attacks. The authors examine a post-training threat scenario in which malicious clients exploit their personalized models to generate adversarial examples targeting other clients. Through experiments across eight PFL methods, the paper identifies substantial cross-client attack transferability and observes a weak correlation between surrogate model accuracy and attack effectiveness. To explain these findings, the authors develop a theoretical analysis identifying five factors influencing adversarial transferability, including model sensitivity, loss landscape geometry, distribution alignment, model fitness differences, and parameter sensitivity. Based on these insights, they propose a defense framework combining stochastic input noise augmentation, input-scaled trace regularization, and parameter sensitivity maximization. Experiments on CIFAR-10, CIFAR-100, and GTSRB demonstrate improvements in clean accuracy and reductions in attack-induced accuracy degradation across multiple PFL methods.

Critical strengths
  1. Relevant security problem in personalized federated learning. The paper investigates cross-client adversarial transferability in personalized federated learning during the post-training inference stage. This is a relevant security concern, as personalized models may remain vulnerable to adversarial examples generated using other clients' models. The observation that surrogate model accuracy is weakly correlated with attack effectiveness is particularly interesting and provides some motivation for further investigation.

  2. Relatively comprehensive empirical evaluation. The authors evaluate adversarial transferability across eight personalized federated learning methods and multiple datasets, including CIFAR-10, CIFAR-100, and GTSRB. The experiments cover different personalization strategies and provide useful evidence that cross-client adversarial transferability is not limited to a single PFL algorithm. The reported defense results also suggest that the proposed approach can mitigate the evaluated transfer attacks while maintaining

Critical weaknesses
  1. Insufficient methodological novelty and lack of a principled framework. Shifting the attack assumption from the training stage to the inference stage is not, by itself, a sufficient novelty. The proposed defense combines stochastic input noise augmentation, input-scaled trace regularization, and parameter sensitivity maximization, but these components are presented as incremental tricks stacked to improve empirical results rather than as a coherent, logically motivated framework.

  2. Questionable theoretical justification of the defense. The theoretical analysis relies on strong convexity assumptions that are difficult to reconcile with the deep neural networks used in the experiments. Moreover, the Hessian trace regularization is not sufficiently justified, since increasing the trace does not necessarily increase the minimum eigenvalue. These issues are consequential because they cast doubt on whether the theoretical analysis actually supports the proposed defense mechanism.

Questions for authors

Beyond the novelty in problem definition, can you clarify the logical and systematic nature of your methodological contributions? That is, you need to show that the method has an internal logical framework, rather than merely being a set of incremental tricks tuned and stacked together to produce the experimental results.

Why is the significance of the problem definition (i.e., moving the attack assumption from the training stage to the inference stage) important enough to warrant a full paper? It is suggested that the above two questions be illustrated with schematic diagrams, comparison tables, and some experiments.

Are you sure that all the optimization techniques you used are truly justified? (For example, the strong convexity assumption and the justification of the trace regularization.)

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.