On Adversarial Robustness of Language Models in Transfer Learning

2024-12-29

AI Summary

While transfer learning improves large language models’ (LLMs) performance on standard NLP tasks, its impact on adversarial robustness remains underexplored and poorly understood. Method: We systematically evaluate multiple architectures—including BERT, RoBERTa, GPT-2, Gemma, and Phi—across the MBIB benchmark suite (covering hate speech, political bias, and gender bias) to assess adversarial robustness under transfer-based fine-tuning. Contribution/Results: We uncover, for the first time, a strong positive correlation between model scale and adversarial robustness—challenging the prevailing assumption that transfer learning inherently enhances robustness. Moreover, we identify intricate couplings among model size, architecture, and adaptation strategy. Empirically, scaling up model parameters mitigates the robustness degradation induced by transfer learning. Our work establishes a reproducible adversarial evaluation framework and provides critical design principles for deploying LLMs that jointly optimize both task performance and security-critical robustness.

Technology Category

• Machine Learning: Adversarial Learning & Robustness
• Natural Language Processing: Safety and Robustness
• Computer Vision: Adversarial Attacks & Robustness

Application Category

• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation
• Security and Privacy: Large-scale security measurements
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs

Abstract

We investigate the adversarial robustness of LLMs in transfer learning scenarios. Through comprehensive experiments on multiple datasets (MBIB Hate Speech, MBIB Political Bias, MBIB Gender Bias) and various model architectures (BERT, RoBERTa, GPT-2, Gemma, Phi), we reveal that transfer learning, while improving standard performance metrics, often leads to increased vulnerability to adversarial attacks. Our findings demonstrate that larger models exhibit greater resilience to this phenomenon, suggesting a complex interplay between model size, architecture, and adaptation methods. Our work highlights the crucial need for considering adversarial robustness in transfer learning scenarios and provides insights into maintaining model security without compromising performance. These findings have significant implications for the development and deployment of LLMs in real-world applications where both performance and robustness are paramount.

Problem

• Language Model
• Robustness
• Adversarial Interference

Innovation

• Large Language Models
• Robustness against Attacks
• Learning Capacity vs Defense Performance