SPARSEBFA: ATTACKING SPARSE DEEP NEURAL NETWORKS WITH THEWORST-CASE BIT FLIPS ON COORDINATES

Deep neural networks (DNNs) are shown to be vulnerable to a few carefully chosen bit flips in their parameters, and bit flip attacks (BFAs) exploit such vulnerability to degrade the performance of DNNs. In this work, we show that DNNs with high sparsity that typically result from weight pruning have a unique source of vulnerability to bit flips when their coordinates of nonzero weights are attacked. We propose SparseBFA, an algorithm that searches for a small number of bits among the coordinates of nonzero weights when the parameters of DNNs are stored using sparse matrix formats. Using SparseBFA, we find that the performance of DNNs drops to the random-guess level by flipping less than 0.00005% (1 in 2 million) of the total bits.