Study on poisoning attacks on CANDECOMP/PARAFAC tensor decomposition for binary tensors

Abstract

Tensor decomposition methods have been widely used to represent a tensor into a compact representation of vectors or set of vectors. These representations should be able to reproduce the original input. Poison attacks occur when we change the input tensor data therefore negatively affecting the performance of succeeding tasks such as classification. In order to investigate how different tensor decomposition algorithms are vulnerable to attacks, we propose two methods of flipping the binary input tensors, namely random attack method and minimum difference method. Simulations show that applying minimum difference attack incurs greater reconstruction error than random attack. We also show that Alternating Poisson Regression algorithm is more vulnerable than Alternating Least Squares Algorithm when minimum difference attack is applied in terms of classifying movie genres.