acceptodds
Withdrawn from ICLR 2027

SpecSIA: Reliable Source Inference in Federated Learning via Adaptive Spectral Filtering

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Federated Learning (FL) enables collaborative model training without directly sharing raw data, yet client-specific updates can still reveal which client contributed a target sample, exposing FL to Source Inference Attack (SIA). Existing passive attacks typically attribute the target to the client model with the lowest prediction loss, which faces two limitations: first, global aggregation propagates target-relevant knowledge across clients, obscuring each client’s original contribution. Second, source attribution can be unstable across diverse federated configurations. Therefore, we propose Spectral Contribution-based Source Inference Attack (SpecSIA), which estimates target-relevant client contributions from current-round updates rather than relying on endpoint prediction loss. By combining normalized contribution attribution with adaptive spectral filtering, SpecSIA improves stability across communication rounds and federated settings. The attack operates using only information available in the current communication round and requires neither local optimization trajectories nor historical client updates. We evaluate source inference from four complementary perspectives: attack effectiveness, temporal reliability, statistical reliability across repeated runs, and cross-configuration consistency. Compared with the conventional loss-based attack, SpecSIA improves mean attack success rate (ASR) by up to 51.88% and reduces temporal variation by up to 54.34%, while maintaining competitive performance across heterogeneous federated configurations.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Federated Learning, Source Inference Attack, Privacy Leakage

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 4 reviews, above 27% of the others.
Reviewer_iYo1OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

SpecSIA scores each client by the inner product between the normalized target gradient and the client's normalized current-round update (RCA, i.e., cosine similarity), after applying a median-calibrated spectral shrinkage to the normalized updates (ASF). The target is attributed to the highest-scoring client. Averaged over each of three datasets, it reports higher ASR and lower Temporal Std than loss-based SIA.

Critical strengths

The method is simple and cheap (one target gradient plus a K×KK\times K solve), and uses the same server view as loss-based SIA, with no shadow models or historical updates. Evaluating SIA beyond mean ASR, in terms of temporal stability and cross-configuration consistency, is a meaningful perspective. RCA (and hence SpecSIA) yields clear gains over loss-based SIA on MNIST and Synthetic.

Critical weaknesses
  1. Flawed motivation. Since ℓ(θt;z)\ell(\theta_t;z) is common to all clients, ranking by the exact round-anchored contribution gives arg⁡max⁡i[ℓ(θt;z)−ℓ(θi,t;z)]=arg⁡min⁡iℓ(θi,t;z)\arg\max_i [\ell(\theta_t;z)-\ell(\theta_{i,t};z)] = \arg\min_i \ell(\theta_{i,t};z), i.e., loss-based SIA. The gain thus stems from linearization or normalization, not from isolating "the contribution introduced by the current client update"; indeed, in Table 1 the gain on CIFAR-10 and Synthetic is near zero at E=1E=1 (where linearization is most accurate) and largest at E=10E=10. Ablations g⊤Δθig^\top\Delta\theta_i and Δℓi/∥Δθi∥\Delta\ell_i/\Vert\Delta\theta_i\Vert are missing.
  2. ASF brings no meaningful benefit. ASF changes ASR by −0.12-0.12 and Temporal Std by −0.04-0.04, far below RCA's +3.68+3.68 / −1.07-1.07, without significance tests. ASR rises monotonically with weaker filtering (larger γ\gamma); the authors concede dominant directions "may still contain useful source-discriminative information", contradicting ASF's premise. The Rank IQR gain (2.00→1.00) rests on coarse integer ranks; per dataset (Table 5), ASF worsens Mean Rank on all three datasets and Rank IQR on CIFAR-10. Yet ASF is the title component.
  3. Limited novelty. RCA is gradient–update cosine similarity (cf. TracIn / influence functions, not discussed). The ASF score equals, up to a positive constant, the ridge-regression coefficients of g^\hat g on {bi}\lbrace b_i\rbrace.
  4. Theory does not support the method. The error bound applies to the unnormalized score, not the normalized score actually used. The scale-invariance result is vacuous because ∑jμj≡1\sum_j\mu_j\equiv1 when ∥bi∥=1\Vert b_i\Vert=1.
  5. Limited experiments. The evaluation covers only FedAvg with K=10K=10, full participation and small models. Additional experiments under broader FL settings are needed.
Questions for authors

Please refer to weakness.

Reviewer_MuuQOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper explores the problem of "source inference" in federated learning, where a passive server attempts to determine which client a known training sample originated from. The proposed method, SpecSIA, performs this inference by comparing the "target gradient under the global model from the previous round" with the "normalized client update from the current round." It employs an adaptive spectral filtering technique to attenuate high-energy components within the client update matrix before ranking the candidate clients. The evaluation covers three datasets and 27 different FedAvg configurations, focusing on the average attack success rate, temporal dynamics, and ranking performance across various configurations. Experimental results demonstrate that SpecSIA significantly outperforms loss-based inference methods, particularly on the MNIST and Synthetic datasets.

Critical strengths
  1. The threat model is clean and the method is inexpensive. It only uses current-cycle information and does not require historical trajectories, auxiliary data, or active intervention; ASF falls in K×K space and its overhead is negligible. 2 The evaluation perspective that "reliability should not be judged by mean ASR alone" is valuable. By reporting metrics such as temporal standard deviation, cross-run LCB, and mean rank (along with rank IQR), the authors demonstrate a rigorous approach, especially given the scope of this subfield (covering 27 configurations across 5 random seeds). 3 The reporting is transparent. LTMIA is stronger on CIFAR-10, θ_0 anchor is better overall but not adopted, and Rank IQR cannot be judged in isolation; these points are all mentioned in the main text.
Critical weaknesses
  1. The motivation behind "round-anchoring" is effectively moot when it comes to ranking; For a fixed round, ℓ(θ_t; z) is the same for all clients, and the loss reduction of argmax is always equal to the endpoint loss of argmin. The ranking differences all come from the first-order approximation plus cosine normalization, while gradient-cosine is the standard practice in gradient-based inference (Nasr et al. 2019 cited in the paper), but no such baseline is used for comparison.
  2. The "ASF" mechanism mentioned in the title proves largely ineffective in the experiments. Fig. 6: After adding ASF, ASR 25.08→24.96, LCB 22.75→22.53, Temporal Std only 2.09→2.05; In Table 6, the larger γ is (the weaker the filtering), the higher the ASR monotonically, and γ→∞ means it degenerates into RCA. The sensitivity analysis itself essentially suggests that "no filtering" yields the best performance.
  3. The scope is overly narrow, and there is a lack of discussion regarding defenses. The experiments are limited to FedAvg with K=10, full client participation, and small CNN architectures, with ASR values ranging from 20% to 35% (compared to a 10% random baseline). The attack relies on the server to see client-by-client updates, and it becomes ineffective when secure aggregation is enabled. The performance under DP noise is also untested. The attack paper does not discuss the most common defenses, so the threat level is unclear.
Questions for authors

please address the weaknesses.

Reviewer_KBR2OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies passive source inference attacks (SIA) in federated learning, aiming to identify which client contributed a target training sample. It argues that reliable SIA must be stable across communication rounds and federated configurations, not merely achieve high average attack success. The authors propose SpecSIA, combining Round-Anchored Contribution Attribution (RCA), which normalizes target-gradient/client-update alignment around the pre-round global model, and Adaptive Spectral Filtering (ASF), which suppresses dominant cross-client spectral directions while preserving client-specific evidence. Contributions include reframing SIA reliability via temporal and cross-configuration consistency, the SpecSIA method, and extensive evaluation. Experiments use CIFAR-10, MNIST, and Synthetic data under 27 FedAvg configurations with varying Dirichlet heterogeneity and local epochs, five seeds, and metrics including mean ASR, temporal standard deviation, 95% LCB, and rank consistency. Results show up to 51.88% ASR improvement and 54.34% temporal-std reduction over loss-based SIA, while remaining competitive with LTMIA.

Critical strengths

1: The paper reframes SIA reliability as a first-class problem, not merely average attack success. The authors explicitly characterize reliable source inference through temporal stability across communication rounds and consistency across federated configurations, and they design metrics—Temporal Std, 95% LCB, Mean Rank, and Rank IQR—around this framing. This is consequential because the motivating experiments show that loss-based SIA can degrade substantially as aggregation progresses and that performance varies with heterogeneity and local training. By making these failure modes central, the paper motivates a genuinely different evaluation standard for passive SIA and supports the claim that SpecSIA’s gains are not just higher ASR but more reliable attribution. This strengthens the paper beyond a narrow method-only contribution.

2: SpecSIA is a principled, lightweight, and empirically well-supported passive attack. RCA converts first-order target-loss reduction into a scale-invariant directional alignment score anchored at the pre-round model, and ASF suppresses dominant cross-client spectral structure using a K$\times$K Gram matrix, keeping the attack practical for K<<d. The component ablation and 27-configuration evaluation support that RCA improves effectiveness and ASF improves temporal stability, while the method remains passive, reference-free, and requires no auxiliary data or training intervention. This is consequential because it makes the contribution both conceptually clean and operationally plausible for a server-side adversary under standard FedAvg, while the broad validation—three datasets, multiple heterogeneity levels, local epochs, and five seeds—substantiates the reliability claims.

Critical weaknesses

1: The paper does not evaluate SpecSIA under differentially private (DP) federated training, which is a central real-world privacy-protection setting. This is consequential because SpecSIA’s core signals—normalized target-gradient alignment and client-specific spectral components—are precisely the kind of fine-grained directional information that DP noise, gradient clipping, and secure aggregation are designed to obscure. The paper claims reliable passive source inference under standard FedAvg, but without DP training it remains unclear whether the attack’s gains survive even moderate privacy budgets, or whether RCA/ASF merely exploit unprotected update structure. Since the threat model is server-side and passive, DP-FL is the most relevant robustness test; omitting it substantially weakens the practical significance and the “reliable” claim.

2: The evaluation relies on relatively simple datasets, and SpecSIA’s advantage is not robust on CIFAR-10, where it underperforms LTMIA. CIFAR-10, MNIST, and a synthetic MLP setting are useful controlled benchmarks, but they do not establish applicability to modern deep models or realistic heterogeneous data. More importantly, on CIFAR-10, Table 1 shows LTMIA attaining higher ASR and 95% LCB than SpecSIA across configurations, while SpecSIA’s absolute ASR remains roughly 26–35%. This contrasts with stronger relative results on MNIST and suggests the method is sensitive to dataset complexity and does not consistently dominate even among passive baselines. Because the paper’s central claim is cross-configuration reliability, the lack of a clear advantage on the most realistic image benchmark—and the absence of larger-scale or modern-architecture validation—materially limits the strength of the empirical case.

Questions for authors
  1. How effective is SpecSIA on DP-FL?
  2. How effective is SpecSIA in attacking more complex datasets, such as ImageNet?
  3. In the case of a larger number of participating clients (>=100), and when only a subset of clients is selected for aggregation in each round, is SpecSIA effective?
  4. Without a clearly defined threat model, can a malicious client carry out SpecSIA?
Reviewer_TbFGOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes RCA, which attributes a target example to the client update most aligned with its loss gradient, and ASF, a spectral filter intended to stabilize attribution across rounds.

Critical strengths

The attack is simple, inexpensive, and evaluated over a reasonably broad grid of data heterogeneity and local-training configurations. The paper is also mostly clear about the threat model and evaluation metric.

Critical weaknesses

My main concern is that the stated motivation for RCA does not survive a simple algebraic check. At a fixed round,

arg⁡max⁡i[ℓ(θt;z)−ℓ(θi,t;z)]=arg⁡min⁡iℓ(θi,t;z).\arg\max_i[\ell(\theta_t;z)-\ell(\theta_{i,t};z)] =\arg\min_i\ell(\theta_{i,t};z).

Thus, introducing the round anchor cannot by itself remove “inherited global knowledge” or change the attribution. RCA differs from exact local loss reduction through first-order linearization and update normalization, making it essentially a current-round gradient-cosine score. The paper does not compare against exact loss reduction, the raw gradient-update inner product, one-sided normalization, or Grad-Cos/TracIn-style baselines, so the source of the reported gain is unresolved. ASF is also weakly supported. In the aggregate result, RCA→RCA+ASF changes ASR from 25.08 to 24.96, LCB from 22.75 to 22.53, and temporal SD from 2.09 to 2.05, while its mean rank worsens. With only five runs, using a normal 1.96 multiplier for the LCB is not well justified. Finally, the evaluation remains restricted to ten fully participating clients, FedAvg, and small models.

Questions for authors

Can the authors provide a matched decomposition among exact loss reduction, raw dot product, one-sided normalization, and cosine normalization? Is the ASF improvement statistically significant under paired bootstrap or a small-sample tt-interval? How does RCA behave with partial participation, larger client populations, secure aggregation, compression, or alternative federated optimizers?

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

Related papers

Discussion (0)

Sign in to comment.