AGRAMPLIFIER: Defending Federated Learning Against Poisoning Attacks Through Local Update Amplification

Federated learning systems are vulnerable to poisoning attacks where participants can sabotage the collaborative training process by manipulating their local data or updates, which is a significant security concern as these systems are designed for privacy-preserving collaboration across multiple parties.

AGRAMPLIFIER improves detection of malicious updates by amplifying distinguishing features of gradient data, making it easier for existing security mechanisms to identify and filter out poisoned contributions—similar to how enhancing contrast in an image makes differences more visible.

The approach works with all existing security mechanisms rather than replacing them, meaning organizations can upgrade their current federated learning defenses without overhauling their entire systems, improving robustness, accuracy, and computational efficiency simultaneously.

Two complementary techniques are used to identify suspicious updates: one organizes gradient data into sections and extracts extreme values (AGRMP), while the other uses explainable AI to identify which features are most activated (AGRXAI), providing multiple ways to spot anomalies.

Testing across seven diverse datasets demonstrates broad applicability, suggesting AGRAMPLIFIER's effectiveness isn't limited to specific industries or data types, making it potentially useful for various real-world federated learning applications from healthcare to finance.