Leaky Language Models: Stealing Architecture and Inference Optimizations via Per-Token Timing

2026-07-22

AI Summary

This work proposes a black-box reverse engineering method that infers architectural parameters and inference optimization strategies of closed-source large language models solely from the timing of remote API responses. By constructing fine-grained models of per-token generation latency on NVIDIA GPUs and integrating architectural space search with speculative decoding detection algorithms, the approach achieves architecture-level probing without requiring access to model weights or internal logs. Experimental results demonstrate that, on Llama-family models, the true architecture is ranked among the top ten hypotheses in over 90% of cases. Furthermore, the method successfully uncovers that Gemini Flash 2.5 employs speculative decoding with a 128K context window, substantially expanding the frontier of black-box model analysis.

Abstract

This work presents LeakyLMs, a set of attacks that leak proprietary model, architecture, and deployment information from production language models. LeakyLMs is the first to demonstrate that key model and deployment details can be inferred using only token generation timing, even when interacting through remote APIs. LeakyLMs introduces two core attacks. The first attack targets inference optimizations and deployment strategies. For example, our attack detects whether a provider uses speculative decoding, a widely deployed inference-time optimization, and further identifies the context length of the draft model used in the pipeline. Our measurements show that Google Gemini Flash 2.5 uses speculative decoding with a draft context window of approximately 128K tokens. The second attack recovers key architectural properties, including the number of transformer layers, hidden dimension size, and number of attention heads. To achieve this, LeakyLMs builds a detailed and accurate model of token-generation timing on modern NVIDIA GPUs, characterizing how latency scales with model configuration and hardware parameters. The attack then performs a search over the architecture space using this timing model. In experiments with Llama models, the near-correct architectural configuration appears in the top-10 guesses more than 90% of the time.

Problem

• language models
• model extraction
• timing attacks
• inference optimization
• model architecture

Innovation

• Leaky Language Models
• speculative decoding
• per-token timing
• model architecture extraction
• side-channel attack