Federated deepfake detector explanations fail under extreme client heterogeneity
The claim
Federated training reduces Grad-CAM faithfulness only when client data heterogeneity is extreme
Published as: “Does Federated Training Degrade Explainability? A Comparative Analysis of Grad-CAM Faithfulness in Centralized vs. Federated Deepfake Detection”
So what
Practitioners can safely deploy federated deepfake detectors with reliable Grad-CAM explanations as long as client data heterogeneity remains moderate; extreme heterogeneity risks misleading explanations despite high accuracy.
The five-minute version
Federated training does not uniformly degrade Grad-CAM explanation faithfulness in deepfake detection; under moderate heterogeneity (α=0.5, 1.0) federated models achieve higher AOPC than the centralized baseline, but under extreme heterogeneity (α=0.1) faithfulness collapses (negative AOPC) and accuracy drops sharply. The paper provides AOPC numbers, statistical tests (Mann-Whitney p=1.0 for α=0.5,1.0, significant degradation only at α=0.1) and a perfect Spearman correlation (ρ=-1.0) linking heterogeneity to degradation. No abstract was provided, so details are drawn from the introduction, results, and conclusion sections.
What undercuts it the paper’s own results
- α = 1.0: mean AOPC = 0.1657 (σ = 0.3482), a margin of +0.0963 over the centralized baseline.
- The one-sided Mann-Whitney U test (H1: central AOPC > federated AOPC) found the centralized model was not superior in faithfulness at α ∈{0.5, 1.0} (p = 1.0 in both cases), with Cohen’s d of -0.13 and -0.29.
Stated limits 1
- For α ∈{0.3, 0.5, 1.0}, Explanation IoU rose clearly between round 5 and round 10, showing that federated and centralized explanations aligned spatially as training advanced. In all three cases the trajectory had plateaued by round 10, which is why we stopped training there under single-GPU compute constraints.
What’s new here 4
- First quantitative comparison of Grad-CAM faithfulness between centralized and federated deepfake detectors.
- Systematic analysis of data heterogeneity using Dirichlet concentration α and its monotonic relationship to explanation degradation.
- Release of an open-source experimental pipeline for federated XAI evaluation on FaceForensics++.
- Identification that moderate heterogeneity can improve explanation faithfulness relative to centralized training.
Artifacts & methods 6
- EfficientNet-B4system
- Federated Averagingprimitive
- Grad-CAMprimitive
- AOPCevaluation
- FaceForensics++dataset
- Dirichlet α heterogeneity partitioningprimitive