Learn to Hide, Learn to See: DRL-Based Charge Manipulation and Attention-MIL Detection for EV Charging Systems

Smart EV charging stations (EVCSs) coordinate charging to satisfy user demand under grid constraints. However, their reliance on networked control creates a cyber-attack surface. In particular, charge-manipulation attacks (CMAs) adjust session parameters (start/end times, requested energy) to reshape aggregate demand while remaining protocol-compliant. Such manipulations degrade forecasting and market operations, accelerate battery aging, and increase stress on distribution assets. Existing defense approaches largely target overt, handcrafted patterns or depend on thresholds and fully supervised labels. These assumptions do not hold in practice as operational datasets rarely provide slot-level annotations, and adaptive, low-amplitude attacks are designed to mimic normal behavior. Therefore, this paper introduces a novel dual-strategy framework. First, we develop a deep reinforcement learning (DRL) agent that acts as an adversary that learns stealthy CMAs (SCMAs) within operational and protocol constraints while optimizing evasion. Second, we develop an intrusion-detection system (IDS) based on sequence-aware multiple-instance learning (MIL). Each charging session is modeled as a bag of time slots; an encoder with attention pooling learns from bag-level (session) labels, provides slot-level attribution, and supports streaming inference for real-time alerts. Evaluations on real charging data demonstrate accurate detection of subtle, temporally localized attacks and consistent improvement over standard anomaly-detection baselines, with higher accuracy and reduced false-alarm rates. We evaluate our approach on a hybrid dataset—benign sessions from ACN-Data. The results show accurate detection of subtle, temporally localized attacks and consistent improvement over standard anomaly-detection baselines with reduced false-alarm rates. We further illustrate the operational and financial consequences of specific SCMA types using market-level analysis under fixed day-ahead bids.