Research

The paper proposes an XSS attack detection model that uses two-stage Abstract Syntax Tree (AST) analysis, first extracting JavaScript from the HTML AST and then identifying malicious fragments in the JavaScript AST, with a Long Short-Term Memory (LSTM) network making the final classification. The authors report 0.991 accuracy and 0.998 F1 on standard XSS samples. Against adversarial XSS samples, their model keeps a detection rate above 0.982, while most existing models drop below 0.880.

This paper proposes FKLM-PDA, a lightweight multidimensional privacy-preserving data aggregation scheme for smart grids that does not rely on a trusted third party. It packs multidimensional smart meter data into a one-dimensional format, replaces the Paillier cryptosystem with random masking and secret-sharing based key separation, and adds fault tolerance and key-leakage resilience. The authors report that performance evaluations show FKLM-PDA outperforms existing schemes in computation and communication.

This paper proposes PPFPS, a privacy-preserving platoon management scheme for flexible platoon splitting in urban freight delivery. It uses an encrypted Manhattan distance method (EMC), combining bloom filters and the Paillier cryptosystem, to match vehicle locations while keeping them private, with reputation ensuring platoon reliability. Simulations report 66.59% to 78.72% lower computation overhead on the TA side, with communication overhead of a similar order of magnitude to existing schemes.

The paper proposes HP-OTP, a challenge-response one-time password scheme that hardens the password without changing existing password-OTP architectures. The server stores only a non-reversible verification credential, and the device's response combines the candidate password, the possession factor, and a random salt. The authors report that comprehensive evaluation shows HP-OTP is secure and efficient, with verification taking less than 5 milliseconds.

Researchers propose a multi-level, cross-domain conditional anonymous authentication scheme for vehicle-to-grid (V2G) systems, aimed at protecting electric vehicles' privacy during cross-domain authentication. The design uses group signatures with a group key generated via the Chinese Remainder Theorem for grid servers, and proxy signatures for vehicles, plus online/offline signatures to improve efficiency. The authors report lower computational, communication, and storage costs than existing schemes and claim security through a formal security model and analysis.

Warning-Graph is a lightweight framework that infers the type of an ongoing APT attack from a limited set of IoCs. It models attack-related infrastructure nodes as a heterogeneous information network and uses heterogeneous graph contrastive learning with two asymmetric graph encoders, plus an information bottleneck loss to reduce noise. Experiments show a 3- to 5-point improvement in identifying APT attack types over baselines while using fewer labeled samples.

The paper introduces Verifiable Data Capsule (VDC), a scheme for controllable data sharing in cloud computing that lets users verify data capsules and state updates. It proposes a Locally Verifiable Chameleon Tag (LVCT) that binds all data capsules to a constant-size tag, a dual-level authentication structure for verifiable updates, and a trust evaluation protocol for judging cloud server credibility. The authors report security analysis and performance evaluation showing practicability.

Researchers propose generating synthetic EEG samples with a conditional denoising diffusion probabilistic model (DDPM) that adds noise, to address scarce EEG data for emotion recognition. On the DEAP dataset, synthetic data improved classification accuracy by up to 5.6%, with similar positive results on the SADT dataset. The approach outperformed traditional generative adversarial network and DDPM-based methods.

This study proposes an AI-based framework that identifies disease-specific volatile organic compounds (VOCs) in exhaled breath for noninvasive early detection. It combines deep learning and reinforcement learning to improve accuracy and reliability on small datasets. The authors name ammonia as especially elevated in chronic kidney disease (CKD).

SFedCA is a credit assignment-based active client selection strategy for spiking federated learning. It replaces random client selection with credits derived from each client's firing intensity before and after local training, which reflects how far local data departs from the global distribution. Experiments on various non-IID scenarios show it outperforms existing state-of-the-art spiking FL methods and needs fewer communication rounds.

The paper introduces Test-time Correction (TTC), an online 3D detection system that corrects errors during inference without retraining. TTC adds an Online Adapter (OA) module to existing 3D detectors, which generates visual prompts from auxiliary feedback such as 2D detection mismatches, road descriptions, or user clicks. The authors report that TTC significantly improves instant error rectification over frozen 3D detectors, including under limited labels, zero-shot settings, and adverse conditions.

Researchers propose a side-channel analysis framework that ensembles neural networks trained under multiple leakage models, using Bagging to combine complementary features from heterogeneous models. The authors use information entropy and Jensen's inequality to prove the ensemble's predictive distribution approximates the true key distribution more closely. They report that the framework improves key recovery efficiency and robustness and generalizes across platforms and adversarial environments.

KGEES is a system for energy-saving demand response in multi-access edge computing (MEC) that addresses location privacy. It applies k-anonymity geo-obfuscation to protect user location and uses a heuristic to set resource allocation under geo-distortion. On a real-world dataset, it outperforms representative approaches by an average of 1.187× in system utility and 1.192× in energy efficiency, and runs 203.5× faster.