Mozes, Maximilian Attila Janos; (2024) Understanding and Guarding against Natural Language Adversarial Examples. Doctoral thesis (Ph.D), UCL (University College London).

Preview

Text mmozes_thesis.pdf - Other Download (2MB) | Preview

Abstract

Despite their success, machine learning models have been shown to be susceptible to adversarial examples: carefully constructed perturbations of model inputs that are intended to lead a model into misclassifying those inputs. While this phenomenon was discovered in the context of computer vision, an increasing body of work focuses on adversarial examples in natural language processing (NLP). This PhD thesis presents an investigation into such adversarial examples in the context of text classification, focusing on studies to characterize them through both computational analyses and behavioral studies. As computational analysis, we present results showing that the effectiveness of adversarial word-level perturbations is due to the replacement of input words with low-frequency synonyms. Based on these insights, we propose an effective detection method for adversarial examples (Study 1). As behavioral analysis, we present (Study 2) a data collection effort comprising human-written word-level adversarial examples, and conduct statistical comparisons between human- and machine-generated adversarial examples with respect to their preservation of sentiment, naturalness, and grammaticality. We find that human- and machine-authored adversarial examples are of similar quality across most comparisons, yet humans can generate adversarial examples with much greater efficiency. In Study 3, we investigate the patterns of human behavior when authoring adversarial examples, and provide “human strategies” for generating adversarial examples that have the potential to advance automated attacks. Study 4 discusses the NLP-related scientific safety and security literature with respect to more recent large language models (LLMs). We provide a taxonomy of existing efforts related to that topic that are categorized into threats arising from the generative capabilities of LLMs, prevention measures developed to safeguard models against misuse, and vulnerabilities stemming from imperfect prevention measures. We conclude the thesis by discussing this work’s contributions and impact on the research community as well as potential future work arising from the obtained insights.

Download activity - last month

Export as JSONXMLCSV

Download activity - last 12 months

Export as XMLJSONCSV

Downloads by country - last 12 months

Export as CSVJSONXML

Archive Staff Only

View Item