Gradient Feature-Based Collaborative Filtering in Verification Federated Learning with Privacy-Preserving

Although federated learning (FL) improves privacy-preserving by updating parameters without collecting original user data, their shared gradients still leak sensitive user information. Existing differential privacy and encryption techniques typically focus on whether the aggregated gradient is correctly processed and verified only, rather than whether each user is honestly trained locally. To address these above issues, we propose a gradient feature-based collaborative filtering scheme in verification federated learning, where the authenticity of user training is verified using the collaborative filtering (CF) method based on gradient features. Compared with single user gradient detection (such as similarity detection of gradient median), our collaborative filtering scheme can provide more comprehensive and efficient user gradient detection by gradient dimensionality reduction. Also, user gradient security is protected by dynamically generating a mask matrix, and the verifiability of the aggregation result is realized by combining dynamic masks. Finally, we perform comprehensive comparisons and experiments by using CNN models on some classical datasets. Experimental results and analysis show that our scheme outperforms other state-of-the-art schemes, demonstrating the effectiveness of our scheme.