acceptodds
Under review as a conference paper at ICLR 2027

AEGIS: A Byzantine-Resilient Federated Learning Protocol for Highly Non-IID Environments

Abstract

Federated Learning (FL) is vulnerable to Byzantine attacks, where compromised clients submit poisoned updates to degrade the global model. Existing robust aggregators force a harsh trade-off: distance-based methods incur an O(k2d) bottleneck and penalize honest statistical drift, while efficient O(kd) coordinate-wise median filters remain blind to directional poisoning. We introduce Aegis, an O(kd) aggregation protocol for high data-skew environments that replaces distance-based filtering with a two-pass dual-metric anomaly detector using adaptive Median Absolute Deviation (MAD) thresholding and cross-round reputation tracking. On CIFAR-10 under a 30% Byzantine threat, Aegis holds accuracy within 10% of the clean baseline across Sign-Flipping, Label-Flipping, and Volume Spam while preserving linear scalability. We also document where statistical filtering breaks down: variance-envelope attacks (ALIE) collapse the protocol to chance-level accuracy, while Sybil collusion degrades it progressively, failing only once cloned identities grow numerous enough to capture the median majority

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Federated Learning, Byzantine Resilience, Model Poisoning, Non-IID Data, Robust Aggregation, ALIE Attack, Median Absolute Deviation

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_chP7OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper introduces an FL method named AEGIS that is designed to reduce the influence of malicious clients while accommodating honest clients with varied data distributions. The method combines median-based filtering, checks on update magnitude and direction, historical reputation, and limits on claimed dataset size to efficiently aggregate updates. However, the submitted manuscript is incomplete, with multiple missing figures.

Critical strengths

Wrong file uploaded with multiple missing figures.

Critical weaknesses

Wrong file uploaded with multiple missing figures.

Questions for authors

Wrong file uploaded with multiple figures missing. Please check the file before submitting.

Reviewer_b5G3OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes AEGIS, an efficient Byzantine-robust aggregation method for federated learning under non-IID data. It combines anomaly detection, robust statistics, and reputation tracking, and evaluates the method against several Byzantine attacks.

It is an imcomplete paper.

Critical strengths

The paper evaluates several attack types, including ALIE and IPM. This makes important failure cases of the proposed defense visible.

Critical weaknesses
  1. AEGIS fails under ALIE, which is explicitly included in the paper's threat model. This directly weakens the main Byzantine-resilience claim.

  2. The submitted version is incomplete. Several figures are missing or shown as file-path placeholders. The AI Use Statement also appears incomplete because the required TODO template remains in the manuscript.

Questions for authors
  1. How should the Byzantine-resilience claim be interpreted given the failure under ALIE?

  2. The AI Use Statement still contains the required TODO template. Is the disclosure incomplete?

Reviewer_ARoeOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes a Byzantine-resilient federated learning aggregation protocol intended for heterogeneous client data. It combines two-pass coordinate-wise median estimation, cosine and Euclidean anomaly scoring, adaptive median absolute deviation filtering, sample-size clipping, and historical reputation. The paper claims linear aggregation complexity and evaluates the method on CIFAR-10 with simulated heterogeneous clients under several poisoning and Sybil attacks. Results show effectiveness against some conventional attacks but substantial failures under ALIE and stronger IPM attacks.

Critical strengths

The problem is relevant. The aggregation procedure is presented explicitly, and avoiding pairwise client comparisons is a useful practical objective. The authors also report substantial failure cases, including ALIE and stronger IPM attacks, which help clarify the method’s limitations.

Critical weaknesses
  1. Figures 1–4 appear as image-file placeholders rather than usable plots. This prevents assessment of the convergence and attack dynamics discussed in the text. The abstract also introduces (O(kd)) without defining its variables, and the notation is reused inconsistently. These issues materially reduce clarity and assessability.

  2. The contribution of the central two-pass procedure is not demonstrated. Multistage robust aggregation has substantial precedent, including Bulyan, which the paper already cites. The authors should explain precisely what distinguishes their construction and demonstrate the resulting benefit. However, Table 4 shows that removing median decontamination slightly improves IPM accuracy, from 63.39% to 63.59%. This ablation does not support the importance of the proposed two-pass mechanism. Reputation and adaptive thresholding also require more complete ablations.

  3. The evaluation is too narrow and lacks relevant comparisons. Experiments use CIFAR-10 with 30 clients and a small range of shard-based partitions. This is insufficient to establish broad effectiveness in highly heterogeneous federated learning. The claimed scalability advantage is not supported by runtime or memory measurements at larger client counts. The baseline set omits relevant developments such as ClippedClustering, Nearest Neighbor Mixing, and methods that explicitly address Byzantine robustness under partial participation. There is also no attack tailored to circumvent AEGIS’s combined filtering and reputation mechanism.

  4. The reported robustness is insufficient for the paper’s claims. In Table 3, AEGIS achieves only 10.00% accuracy under ALIE, compared with 66.78% for FedAvg and 71.15% for FoolsGold. Under IPM, AEGIS also underperforms both methods. Table 2 further shows accuracy falling to 17.69% as the IPM magnitude increases. These results indicate that the proposed filtering can substantially amplify an established attack. Reporting this failure is valuable, but it does not demonstrate the proposed method's advantage.

Questions for authors
  1. What improvement does the second median pass provide over a single-pass version across the evaluated attacks and heterogeneous settings?

  2. Can the authors provide repeated-run results with uncertainty estimates, together with runtime measurements at substantially larger client counts?

  3. How does AEGIS compare with stronger heterogeneous-data defenses and attacks adapted to its scoring rule?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.