PyTorch backward pass exploited to run Turing-complete malware

Researchers have discovered a method to execute Turing-complete programs within the backward pass of PyTorch neural networks. By crafting a loss function that mathematically evaluates to zero, attackers can introduce arbitrary code execution, such as cellular automata like Conway's Game of Life, into the model's training process. This exploit leverages discrepancies between IEEE 754 floating-point arithmetic and PyTorch's automatic differentiation engine, allowing malicious payloads to be embedded without altering the model's reported loss or requiring direct modification of the base model checkpoint. AI

IMPACT This exploit highlights a novel attack vector in ML frameworks, potentially impacting the security of model fine-tuning and deployment.

RANK_REASON Technical paper detailing a novel exploit in a popular ML framework. [lever_c_demoted from research: ic=1 ai=1.0]

Read on Lobsters — AI tag →

• AutoModelForCausalLM
• Conway's Game of Life
• Elon Litman
• Hugging Face
• IEEE 754
• llama
• Meta*
• Meta-Llama-3-8B
• PyTorch
• UltraLoss
• UltraLossRegularizer

• other

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →