Federated deepfake Grad-CAM faithfulness collapses at extreme heterogeneity
Original title: Does Federated Training Degrade Explainability? A Comparative Analysis of Grad-CAM Faithfulness in Centralized vs. Federated Deepfake Detection
An EfficientNet-B4 on FaceForensics++ matches a centralized baseline until client data is extremely non-IID, when explanations turn misleading.
Researchers compare Grad-CAM faithfulness for an EfficientNet-B4 deepfake detector trained centrally and across five federated clients on FaceForensics++. Near-IID and moderate heterogeneity match or exceed the centralized baseline. At extreme heterogeneity, explanations become spatially incoherent and, on average, causally misleading. They treat this as a heterogeneity threshold rather than a general cost of federation, and release code and data partitions.
- The detector is EfficientNet-B4 evaluated on FaceForensics++.
- Training spans five federated clients with varying data heterogeneity.
- Near-IID and moderate heterogeneity match or beat centralized faithfulness.
- Extreme heterogeneity makes explanations spatially incoherent and causally misleading.
- Code and data partitions are released for reproducibility.
- Products
- FaceForensics++
- AI models
- EfficientNet-B4
From the original article
1 min read at arxiv.orgAutomated detection has become the essential safeguard against realistic synthetic media, but deploying it at scale requires training on sensitive data from multiple organizations, putting privacy and the transparency demanded by forensic review in direct conflict.
By Muhammad Uzair Shahid · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.