Detecting Malicious Gradients from Asynchronous SGD on Variational Autoencoder
Zhipin Gu, Yuexiang Yang, Heyuan Shi · 2021
In asynchronous distributed learning, the parameter server updates the global model as soon as a new gradient is received from any device. The asynchronous systems are designed to address the existence of lagging devices which is inevitable due to device heterogeneity and network unreliability. However, the lack of synchrony incurs additional noise and makes detecting and defending against malicious model gradients a challenging task. Unlike existing works that struggle to design robust methods to tolerate untargeted model poisoning gradients, the paper considers detecting and removing targeted model poisoning gradients from the normal asynchronous training process. This paper proposes Asynvae, a robust distributed asynchronous learning framework where the parameter server uses variational autoencoder to detect and exclude malicious gradients. Since the reconstruction error of malicious updates is much larger than that of benign ones, it can be used as an anomaly score. We formulate a threshold of reconstruction error to differentiate malicious updates from normal ones based on this idea. Asynvae is tested with extensive experiments on distributed learning benchmarks, showing a competitive performance over existing distributed learning methods under untargeted model poisoning attack, targeted model poisoning attack and lagging attack.