Identifying and Mitigating the Security Risks of Generative AI
arXiv:2308.14840 · doi:10.1561/3300000041
Abstract
Every major technical invention resurfaces the dual-use dilemma -- the new technology has the potential to be used for good as well as for harm. Generative AI (GenAI) techniques, such as large language models (LLMs) and diffusion models, have shown remarkable capabilities (e.g., in-context learning, code-completion, and text-to-image generation and editing). However, GenAI can be used just as well by attackers to generate new attacks and increase the velocity and efficacy of existing attacks. This paper reports the findings of a workshop held at Google (co-organized by Stanford University and the University of Wisconsin-Madison) on the dual-use dilemma posed by GenAI. This paper is not meant to be comprehensive, but is rather an attempt to synthesize some of the interesting findings from the workshop. We discuss short-term and long-term goals for the community on this topic. We hope this paper provides both a launching point for a discussion on this important topic as well as interesting problems that the research community can work to address.
References in corpus (36)
- Training language models to follow instructions with human feedback
- BERTScore: Evaluating Text Generation with BERT
- Emergent Abilities of Large Language Models
- Artificial Intelligence, Values and Alignment
- ReAct: Synergizing Reasoning and Acting in Language Models
- Constitutional AI: Harmlessness from AI Feedback
- Release Strategies and the Social Impacts of Language Models
- Self-Refine: Iterative Refinement with Self-Feedback
- Leveraging Frequency Analysis for Deep Fake Image Recognition
- Universal and Transferable Adversarial Attacks on Aligned Language Models
- TweepFake: about Detecting Deepfake Tweets
- Unmasking DeepFakes with simple Features
- Can AI-Generated Text be Reliably Detected?
- The Curse of Recursion: Training on Generated Data Makes Models Forget
- DetectGPT: Zero-Shot Machine-Generated Text Detection using Probability Curvature
- ChemCrow: Augmenting large-language models with chemistry tools
- Large Language Models for Code: Security Hardening and Adversarial Testing
- Red Teaming Language Models to Reduce Harms: Methods, Scaling Behaviors, and Lessons Learned
- A Watermark for Large Language Models
- Recipes for Safety in Open-domain Chatbots
- On the Robustness of ChatGPT: An Adversarial and Out-of-distribution Perspective
- Paraphrasing evades detectors of AI-generated text, but retrieval is an effective defense
- Teaching language models to support answers with verified quotes
- PromptRobust: Towards Evaluating the Robustness of Large Language Models on Adversarial Prompts
- Exploiting Programmatic Behavior of LLMs: Dual-Use Through Standard Security Attacks
- Are aligned neural networks adversarially aligned?
- On the Reliability of Watermarks for Large Language Models
- LLM Censorship: A Machine Learning Challenge or a Computer Security Problem?
- Robust Distortion-free Watermarks for Language Models
- Provable Robust Watermarking for AI-Generated Text
- Detecting Language Model Attacks with Perplexity
- Data Feedback Loops: Model-driven Amplification of Dataset Biases
- Undetectable Watermarks for Language Models
- Rethinking Model Evaluation as Narrowing the Socio-Technical Gap
- PTW: Pivotal Tuning Watermarking for Pre-Trained Image Generators
- Coarse-Tuning Models of Code with Reinforcement Learning Feedback