Repurposing Backdoors for Good: Ephemeral Intrinsic Proofs for Verifiable Aggregation in Cross-silo Federated Learning

Researchers propose a novel lightweight architecture for verifiable aggregation in federated learning that uses backdoor injection as intrinsic proofs instead of expensive cryptographic methods. The approach achieves over 1000x speedup compared to traditional cryptographic baselines while maintaining high detection rates against malicious servers.

• →New method replaces heavyweight cryptographic proofs with intrinsic proofs embedded directly into model parameters using backdoor injection.
• →Achieves over 1000x performance improvement on ResNet-18 compared to existing cryptographic verification methods.
• →Uses catastrophic forgetting to make verification signals ephemeral, preserving final model utility while enabling immediate verification.
• →Demonstrates high detection probabilities against malicious servers on standard datasets including SVHN, CIFAR-10, and CIFAR-100.
• →Provides a scalable solution for cross-silo federated learning that maintains client anonymity without requiring trusted third parties.