Manjinder Singh

Manjinder Singh, Alexander E. I. Brownlee, David Cairns

Metaheuristic search algorithms look for solutions that either maximise or minimise a set of objectives, such as cost or performance. However most real-world optimisation problems consist of nonlinear problems with complex constraints and conflicting objectives. The process by which a GA arrives at a solution remains largely unexplained to the end-user. A poorly understood solution will dent the confidence a user has in the arrived at solution. We propose that investigation of the variables that strongly influence solution quality and their relationship would be a step toward providing an explanation of the near-optimal solution presented by a metaheuristic. Through the use of four benchmark problems we use the population data generated by a Genetic Algorithm (GA) to train a surrogate model, and investigate the learning of the search space by the surrogate model. We compare what the surrogate has learned after being trained on population data generated after the first generation and contrast this with a surrogate model trained on the population data from all generations. We show that the surrogate model picks out key characteristics of the problem as it is trained on population data from each generation. Through mining the surrogate model we can build a picture of the learning process of a GA, and thus an explanation of the solution presented by the GA. The aim being to build trust and confidence in the end-user about the solution presented by the GA, and encourage adoption of the model.

Manjinder Singh, Alexander E. I. Brownlee, Mohamed Elawady

Deep learning models have achieved impressive performance across various fields but remain vulnerable to adversarial inputs, particularly in NLP, where such attacks can have significant real-world consequences. Adversarial attacks often involve small, semantically similar token replacements to fool NLP models, and recent methods have become more precise by targeting specific vulnerable words, often by exploiting some level of access to the model's internal structure. This paper proposes GAversary, a hybrid Genetic Algorithm (GA) to generate adversarial attacks on natural language models. The GA is able to treat the target model as a black box, requiring only the logit value output by the model to guide the search. GAversary differs from GAs previously proposed for this problem by using GloVe embeddings to propose word replacements (the mutation operator) to improve the semantic similarity of the adversarial examples. GAversary is applied to several benchmark data sets and well-known target models. GAversary is able to substantially reduce the target model's accuracy on test data compared to the BAE and A2T attacks compared against (in the best case, reducing a 76.8% accuracy to 5.8%, compared to BAE's 27.6%). The trade-off is that GAversary perturbs just under twice as many words as the other two methods, with a slightly lower semantic similarity to the original text and around a 5% increase in run-time.