Differentiable Language Model Adversarial Attacks on Categorical Sequence Classifiers

Abstract

An adversarial attack paradigm explores various scenarios for the vulnerability of deep learning models: minor changes of the input can force a model failure. Most of the state of the art frameworks focus on adversarial attacks for images and other structured model inputs, but not for categorical sequences models. Successful attacks on classifiers of categorical sequences are challenging because the model input is tokens from finite sets, so a classifier score is non-differentiable with respect to inputs, and gradient-based attacks are not applicable. Common approaches deal with this problem working at a token level, while the discrete optimization problem at hand requires a lot of resources to solve. We instead use a fine-tuning of a language model for adversarial attacks as a generator of adversarial examples. To optimize the model, we define a differentiable loss function that depends on a surrogate classifier score and on a deep learning model that evaluates approximate edit distance. So, we control both the adversability of a generated sequence and its similarity to the initial sequence. As a result, we obtain semantically better samples. Moreover, they are resistant to adversarial training and adversarial detectors. Our model works for diverse datasets on bank transactions, electronic health records, and NLP datasets.Preprint. Under review.

Adversarial Attacks on Deep Models for Financial Transaction Records

“…Adversarial attacks and defences from them are of crucial importance in financial applications [28]. The literature of adversarial attacks on transaction records includes [8,9]. However, the authors do not consider the peculiarities of transaction records data and apply general approaches to adversarial attacks on discrete sequence data.…”

“…However, FGSM-based attacks provide better performance scores than SamplingFool-based ones due to the random search for adversarial examples in the second case. It can be useful to unite these approaches to create a generative model that can generate sequences of transaction records that are both realistic and adversarial [8]. Also, more realistic concatenation of tokens to the end of a sequence results in lower performance scores.…”