Beyond Gradient and Priors in Privacy Attacks: Leveraging Pooler Layer Inputs of Language Models in Federated Learning

Federated learning (FL) emphasizes decentralized training by storing data locally and transmitting only model updates, underlining user privacy. However, a line of work on privacy attacks undermines user privacy by extracting sensitive data from large language models during FL.Yet, these attack techniques face distinct hurdles: some work chiefly with limited batch sizes (e.g., batch size of 1), and others can be easily defended or are transparently detectable. This paper introduces an innovative approach that is challenging to detect and defend, significantly enhancing the recovery rate of text in various batch-size settings. Building on fundamental gradient matching and domain prior knowledge, we enhance the recovery by tapping into the input of the Pooler layer of language models, offering additional feature-level guidance that effectively assists optimization-based attacks. We benchmark our method using text classification tasks on datasets such as CoLA, SST, and Rotten Tomatoes. Across different batch sizes and models, our approach consistently outperforms previous state-of-the-art results.