acceptodds
Under review as a conference paper at ICLR 2027

APTI-FL: Adversarial Perturbation Guided Trigger Inversion for Backdoor Detection in Federated Learning

Abstract

Federated learning limits the server's visibility into local training, allowing compromised clients to inject backdoors into a shared model. Existing approaches struggle to detect such malicious clients when non-IID data makes benign updates closely resemble malicious ones, or when attackers constitute the majority of participating clients. In our preliminary study, we discover a distinctive property of backdoored models: although adversarial perturbations substantially disrupt normal classification paths, a triggered input remains likely to be classified into the specified target even after the same perturbation is added. In other words, the backdoor path exhibits target retention under adversarial perturbations. Motivated by this observation, we propose APTI-FL, an adversarial attack guided federated backdoor defense. APTI-FL first generates untargeted adversarial perturbations from trusted server data and uses discrete and continuous region inversion paths to recover potential triggers with different spatial structures. It then determines a reliable attack target through multistage validation of candidate trigger paths and target correction across rounds. Finally, it combines multidimensional client behavior with historical evidence to identify clients that exhibit backdoor responses and blocks their contributions before aggregation. Experiments on five benchmark datasets against five representative defenses show that APTI-FL substantially reduces attack success rates while maintaining main task accuracy.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Federated Learning, Backdoor Attack, Non-IID, Adversarial Attack, Trigger Inversion.

Reviews

Among:
Shape:
1234this paper, Bottom 9%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.50 over 2 reviews, above 9% of the others.
Reviewer_ZanDOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

APTI-FL detects federated backdoors using a small trusted server dataset. It searches for triggers that cause both natural and adversarially perturbed samples to predict the same target. It then checks which clients strengthen this response and excludes them from aggregation, using client history to stabilize the decisions. Experiments on five datasets show substantial reductions in attack success under CBA and DBA, including settings with a malicious-client majority.

Critical strengths
  1. The detection idea is reasonable. Testing what a client adds beyond the incoming global model can help distinguish new backdoor contributions from an inherited backdoor. Matched control inputs also help account for ordinary class preference.
  2. The results are promising within the tested settings. The evaluation covers five datasets and both centralized and distributed patch attacks, and shows that the full method can remain effective when malicious clients outnumber benign clients.
Critical weaknesses
  1. The baseline comparisons and ablations do not sufficiently establish the method’s added benefit. The main tables use 50% and 60% malicious-client ratios, which do not ensure a benign majority among participating clients. Several baselines depend on this condition; for example, FLAME explicitly assumes fewer than 50% malicious clients. Failure when this assumption is violated cannot by itself establish general superiority. The authors should state each baseline’s assumptions and add comparisons within its supported operating range. APTI-FL also uses 250 trusted server samples, but does not compare with FLTrust, which uses the same type of resource and is already cited. A comparison using the same trusted data is needed to judge whether the more complex design provides an additional benefit. Trigger inversion has been studied in prior work, but the paper does not sufficiently explain its improvements over existing methods. It should discuss representative inversion methods and clarify what adversarial guidance adds and which difficulty in the federated setting it addresses. The paper also does not directly test its key component: adversarial guidance. Fig.6 varies loss weights, but does not isolate the benefit of this component. Keeping the client-detection pipeline fixed, the authors could compare with an existing inversion method, remove PGD, or replace it with random perturbations. Such comparisons would help establish the added benefit of the proposed design.
  2. The experiments leave important operating conditions and costs untested. The malicious-client sweep begins at 40%, although Fig.4 already shows deterioration at that ratio in some cases. Results at 10–30% and with no attackers are important: weaker backdoor signals may be harder to invert, while benign-only training reveals unnecessary client exclusion. The paper also fixes the trusted dataset at 250 samples and does not test smaller or class-biased sets. The minimum Dirichlet parameter is 0.35; stronger label skew and actual client class distributions should be examined. Trigger inversion runs every round. The stated CIFAR-100 budget implies 9,000 discrete inversion optimizer steps per round, before other stages. Without measured runtime, it is difficult to judge whether the benefit justifies this recurring computation.
Questions for authors
  1. What attacker-fraction assumptions do the baselines make, and do they hold for the participating clients in these experiments? Can you add comparisons within their supported ranges and with FLTrust using the same trusted samples?
  2. What does the trigger-inversion design add over prior work? With the client-detection pipeline held fixed, how do results compare when using an existing inversion method, removing PGD, or replacing PGD with random perturbations?
  3. What are the results at 0–30% malicious clients, with smaller or class-biased trusted datasets, and under stronger label skew? Please include client FPR and clean accuracy.
  4. What are the measured per-round server time and peak memory, compared with FedAvg and a trusted-data baseline? Please state the hardware and participating-client count.
Reviewer_w3YKOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes a backdoor defense method for federated learning. The defending server generates adversarial perturbations on trusted data and performs trigger inversion in both discrete and continuous regions to search for candidate triggers that consistently induce the same target class in the presence of perturbations. Using the candidate triggers, the server compares the responses of individual client models, the global model from the previous round, and leave-one-out aggregated models. It then combines these responses with each client’s contribution to the benign task and historical evidence across rounds to compute scores for identifying malicious clients. Finally, the server sets the aggregation weights of suspicious clients to zero, excluding them from the global update to defend against backdoor attacks.

Critical strengths
  1. This paper proposes an interesting phenomenon that clean inputs and backdoored inputs exhibit different levels of stability when exposed to adversarial perturbations. Clean inputs are more easily affected by adversarial perturbations, whereas backdoors can maintain a stable target-class activation effect.
  2. Trigger inversion considers both discrete and continuous regions. The strategy for filtering potentially malicious clients takes multiple factors into account, including historical records, the performance of clients in the same round, the previous-round global model, and leave-one-out results. This reduces the likelihood of misclassifying benign clients in non-IID settings.
  3. The experiments cover multiple datasets and different proportions of malicious client participation. They also evaluate performance under various Dirichlet-based data partitions, demonstrating excellent results.
Critical weaknesses

The paper evaluates its defense only against two backdoor attacks, CBA (2019) and DBA (2020), and does not provide validation against more recent backdoor attack methods. Moreover, both CBA and DBA use fixed triggers, although DBA employs distributed triggers, the paper does not assess the defense against adaptive backdoor attacks. These limitations raise doubts about the claimed effectiveness of the proposed method. In particular, it remains unclear whether APTI-FL can maintain its defense performance when the attacker employs adaptive or weak-signal backdoor attacks. 2. The motivation of this paper is based on the observation that clean models and backdoored models exhibit different levels of stability under adversarial perturbations. Although Figure 2 presents corresponding experimental results, these findings are empirical and may not be universally applicable. The perturbation magnitude can substantially affect this conclusion, and whether this phenomenon exists across different types of backdoor attacks requires further investigation. 3. Please report the detailed computational overhead of the proposed method and the baselines. The computational cost of this method increases with the number of classes, the number of clients, and the model size. Moreover, the model used in the current experiments is relatively small and may not adequately reflect the real-world federated learning scenarios described by the authors.

Questions for authors

See weaknesses.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.