RelayGuard: Budget-aware burst shaping against traffic analysis in dual-hop privacy relays

Dual-hop privacy relay systems enhance user privacy by separating client identity from destination visibility through a split-trust architecture. However, despite this protection model, such systems remain vulnerable to traffic analysis attacks that exploit packet timing, burst patterns, and flow-level metadata to infer visited websites or correlate sessions. Existing defenses against traffic analysis are often too costly for low-latency privacy relays because they can introduce substantial bandwidth and latency overhead. In this paper, we present RelayGuard , a lightweight relay-side defense for dual-hop privacy relays based on budget-aware burst shaping at the ingress relay. RelayGuard transforms observed traffic into canonical burst patterns using a greedy scheduling strategy with budget correction, while encouraging a near-anonymity property through k-anonymous burst buckets , where each bucket is shared by at least k website or page classes. The defense operates on four burst-level features—direction, burst size, inter-burst gap, and burst duration—and applies three practical actions: adaptive relay-layer padding, small delay injection, and short-window micro-batching. We formulate the defense objective as a joint optimization of canonicalization fidelity, anonymity preservation, and overhead cost, and solve it using an efficient online heuristic suitable for low-latency ingress-side deployment. We evaluate RelayGuard in a controlled emulated dual-hop relay testbed under defense-aware website fingerprinting and session linkage attacks. Experimental results show that RelayGuard reduces closed-world website fingerprinting accuracy from 84.22% to 22.45% for a classical attacker and from 93.58% to 28.90% for a deep learning-based attacker. It also lowers page-family-level session linkage accuracy to 38.50% and 46.75% for the two attackers, respectively, while maintaining practical shaping overhead. Additional comparisons against recent shaping-inspired baselines show that RelayGuard provides a stronger privacy–overhead trade-off in the evaluated relay setting. These results suggest that budget-aware burst shaping is a promising direction for strengthening traffic analysis resistance in dual-hop privacy relay architectures.