A Large-Scale Benchmark and Risk Assessment of Traffic Analysis Attacks on Cloud LLM Services

Burst features alone achieve 94.17% model-fingerprinting accuracy versus 95.67% for all 168 features; removing burst features drops accuracy by 9.33 percentage points, while removing packet-size or timing alone causes near-zero degradation. Information-theoretic analysis confirms burst-byte features leak 3.02 of the maximum 3.32 bits of model identity, versus only ~0.23 bits from aggregate volume/direction.

§V.A.2, Table I detection traffic-shapeml-classifier generic

Prompt reformulation (wrapping prompts with semantics-preserving instructions) reduces model-fingerprinting accuracy from 93.2% to 69.5% (−23.7 pp), and temperature randomization reduces it by up to 18.2 pp, but accuracy remains far above the 10% random baseline in all tested conditions. Because the attacks exploit correlated signals across packet sizes, timing, burst rates, and response progression simultaneously, the paper concludes that defenses must reshape the entire traffic profile rather than mask any single feature family.

§V.C, §VI.A, Figure 9(a) evaluation traffic-shapeml-classifier generic

Using only encrypted packet metadata — packet sizes, directions, timing, and burst structure — a passive on-path observer achieves 97.7% balanced accuracy fingerprinting 10 LLMs, dropping only to 92.2% when the candidate set scales to 20 models. Two post-trained variants sharing the same Qwen3-14B backbone remain separable at 100% and 96.7% recall, showing that post-training-induced behavior changes the traffic fingerprint independently of architecture.

§V.A.2, Figure 9(b) detection traffic-shapeml-classifier generic

Multi-agent task fingerprinting achieves up to 90.7% balanced accuracy using all 247 features; observing encrypted traffic from a single agent yields 66.5% accuracy under graph topology and 78.0% under star topology, rising to 84.3% and 85.9% with up to four agents visible. Cross-topology accuracy drops from 83% to 52% but remains well above the 10% chance level, indicating that topology changes reduce but do not eliminate task leakage.

§V.B, Figure 8 detection traffic-shapeml-classifier generic

Using only LLM-to-user response packets (22 features), a classifier identifies the semantic category of a user's prompt — among six classes including code generation, mathematical reasoning, and adversarial requests — with 67.37% mean accuracy versus a 16.67% random baseline. Response-only accuracy exceeds full-traffic accuracy (65.10%), and packet-size and timing features contribute the most unique information (8.34 pp and 7.26 pp drop when each is removed); the leakage originates from generation behavior, not prompt length.

§V.A.3, Table II–III detection traffic-shapeml-classifier generic