acceptodds
Under review as a conference paper at ICLR 2027

SCARFL: Single-Ciphertext Privacy-Preserving Byzantine-Robust Federated Learning

Abstract

Privacy-preserving Byzantine-robust federated learning requires computing geometric statistics of client updates without exposing the updates themselves. Existing baselines like AegisFL achieve this by extracting encrypted scalars from two coefficient-packed ciphertexts per client, which necessitates an interactive consistency check. In this work, we reveal that the complementary packing is mathematically exactly the inverse-index Galois automorphism of the primary packing. Leveraging this insight, we propose SCARFL, which eliminates the redundant client upload by deriving the second ciphertext view directly at the aggregation server while strictly preserving the encrypted scalar interface. To demonstrate the practical viability of this retained interface, we instantiate it with GHARC, an aggregation rule that discovers update clusters and accepts them only when historical-reference evidence passes a security guard. For 30 clients, our Single-CT protocol halves recurring client uploads and reduces total round traffic by 43.0% on HAR and 45.7% on MNIST. Mean round time falls by 11.0% and 9.2%, respectively, with a negligible maximum inner-product error of . Under evaluated untargeted, label-flipping, and scaling attacks, the GHARC instantiation maintains stable clean accuracy and suppresses attack success rates with up to 14 malicious clients. Comprehensive ablations validate the roles of its clustering, clipping, ranking, and adaptive core expansion mechanisms.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Federated Learning (FL), Privacy Preservation, Robust Aggregation, Homomorphic Encryption, Communication Efficiency

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_PDuCOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

SCARFL simplifies AegisFL by letting the server derive the second ciphertext view from the first using a Galois automorphism. This removes a duplicate client upload and the associated consistency exchange while keeping the existing encrypted statistics interface. The paper also proposes GHARC, a robust aggregation rule based on historical-reference similarity, clustering, clipping, and conditional core expansion. Experiments with 30 clients on MNIST and HAR report 43.0-45.7% less round traffic and 9.2-11.0% less round processing time. Robustness is evaluated against random updates, label flipping, and a triggered scaling attack.

Critical strengths
  1. The communication savings come from a well-controlled comparison. Single-CT and Dual-CT use the same cryptographic parameters and downstream operations. The measurements include the extra Galois computation and its numerical error. This makes the reported protocol savings convincing.

  2. GHARC works within the available encrypted operations. Its decisions require only norms and inner products, so it fits the retained protocol interface. The ablations also show that clustering, clipping, ranking, and core expansion have distinct effects on attack success and training progress.

Critical weaknesses
  1. The main protocol contribution is fairly narrow. The main change is to derive AegisFL's second ciphertext view at the server using a standard Galois automorphism. The encrypted statistics interface, key conversion, and overall protocol structure are largely retained. The communication savings are useful, but this is a fairly narrow optimization.

  2. The two-server trust assumption limits practical deployment. Privacy relies on the aggregation server and the key-holding server never colluding. This is a strong requirement, if they collude, they can combine the ciphertexts and the secret key to recover individual client updates.

  3. The evaluation relies on simple datasets and basic attack settings. The experiments use MNIST and HAR with 30 clients and test Gaussian random updates, label flipping, and a fixed scaling attack. These are established baseline attacks, but they are not enough to assess a new Byzantine defense.

Questions for authors
  1. Can SCARFL be adapted to a single-server architecture while preserving update privacy and Byzantine robustness?

  2. Could you evaluate the method on more challenging datasets and larger client populations, including stronger coordinated or adaptive attacks that take GHARC's reference and selection rules into account?

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

The paper proposes using second ciphertext deriving to avoid interactive consistency checking. The solution is tested on small dataset and model setup such as MNIST and unnamed "lightweight neural models".

Critical strengths

This work is designed to remove the consistency check in FL anomaly detection, which is from a well-described motivation.

Critical weaknesses
  1. The author agrees that the evaluation only covers two lightweight datasets, it is unclear to me from the presented results whether this work can generalize.
  2. The novelty claim seems to be overstated. For example, the paper states "we reveal that the complementary packing is mathematically exactly the inverse-index Galois automorphism", but this is rather standard knowledge.
  3. The paper does not address practical concerns like key generation and key switching costs in this design.
Questions for authors

The paper doesnt state the model details in your experiment setup? What do you use?

Reviewer_AF6POpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes SCARFL method, which is motivated by AegisFL (published in ICML 2024). It designs Single-CT, removing a second upload and the associated consistency exchange. It also provides GHARC, a new robust aggregation rule using authorized norms and inner products to select a small core based on historical-reference similarity. It conducts experiments using two datasets (HAR and MNIST) with 30 clients to show the advantages of SCARFL.

Critical strengths
  1. The Single-CT is elegant with useful cryptographic insight. Theorem B.1 is simple, mathematically exact at the plaintext encoding layer, and immediately actionable. The experiments also show the effectiveness of Single-CT.
  2. Comprehensive ablation study of GHARC shows that each component is needed against the scaling attack.
Critical weaknesses
  1. The experiment is limited. Evaluated only on two small tasks (MNIST and HAR, model sizes 72K and 101K parameters) against three basic attacks (Gaussian noise, label flipping, 5x scaling). Well-known model-poisoning attacks, e.g., ALIE, Min-Max, and Min-Sum, are not evaluated.
  2. GHARC is a heuristic with many hand-set constants and no guarantees. Unlike the clear Single-CT insight, GHARC stacks numerous heuristics and hardcoded constants without sensitivity analysis or a selection method; for example, a0=7a_0=7, Δa=5\Delta a = 5, guard threshold 0.2, ranking weights 0.75 / 0.15 / 0.10, b_c = 0.02, etc.
  3. The Single-CT design is correct, elegant, and well measured, but it is an elementary, known algebraic relation. The cryptographic contribution is incremental.
Questions for authors
  1. How does GHARC perform under ALIE, Min-Max / Min-Sum, and an adaptive attacker who knows GHARC?
  2. Are the improvements over competing aggregation rules consistent across independent seeds and stronger non-IID settings?
  3. What happens with larger models, partial participation, and client dropout?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.