POSTER: Robust Encrypted Traffic Classification under Cross-Scenario Shift
Abstract
Encrypted traffic classification is crucial for network monitoring when payload inspection is unavailable. However, many payload-agnostic classifiers degrade after deployment across different network scenarios, because packet-size and timing statistics shift while target-domain labels are unavailable. We present ShiftGuard, a compact framework for cross-scenario encrypted traffic classification. ShiftGuard learns domain-invariant per-flow representations from labeled source traffic, then performs cautious target-side graph refinement over unlabeled flows. It combines embedding similarity with short-window endpoint co-occurrence to construct a sparse temporal context graph, and applies residual graph-attention correction so that contextual evidence improves uncertain predictions without aggressively propagating errors. Experiments on CrossNet2021 show that ShiftGuard consistently outperforms representative baselines under cross-domain transfer, and an ablation study confirms that each design component contributes to overall robustness.