TARL: Multi-task adaptive representation learning for secure encrypted traffic analysis

The growth of encryption technology has made encrypted traffic analysis both increasingly critical and inherently challenging. Tasks such as intrusion detection, application classification, and encrypted web fingerprinting are vital to network security, yet are often studied in isolation using single-task models. This fragmented approach limits the ability to capture shared behavioral patterns and undermines generalization across tasks. In this paper, we propose TARL, a novel task-adaptive representation learning framework for secure encrypted traffic analysis. TARL combines a shared memory module for task-agnostic representation learning with a task-specific fusion mechanism for downstream adaptation. To support pretraining under encryption constraints, we introduce two self-supervised objectives: masked feature modeling and service-type prediction. We evaluate TARL on four real-world encrypted traffic tasks. Results show that TARL consistently outperforms single-task and multi-task baselines, achieving strong generalization and per-task accuracy. Ablation studies further validate the complementary design of shared and task-specific modules.