Multivariate time series (MTS) forecasting models are vulnerable to training-stage backdoor attacks, where a small number of poisoned windows can make the model predict an attacker-specified future trajectory when a trigger appears. Existing forecasting backdoors mainly generate and place triggers through time-domain or prediction-error heuristics, leaving trigger synthesis, poisoning location, and target trajectory only loosely coupled. This mismatch becomes more pronounced under short-trigger constraints, where the injected segment must be locally compatible with the historical window while remaining effective for steering future predictions. We propose IFBA, a frequency-aware backdoor attack for MTS forecasting. IFBA generates input-aware triggers from the local pre-trigger context, selects attacked variables and poisoning timestamps through a low-frequency spectral-consistency score, and optimizes the trigger generator with a frequency-domain alignment loss. The scheduling score and the generator objective are built on the same magnitude-phase spectral representation, so that trigger synthesis and placement are coordinated by local spectral compatibility rather than treated as separate heuristics. Experiments on five public MTS benchmarks and three forecasting architectures show that IFBA achieves stronger attack effectiveness than existing forecasting backdoor methods while maintaining comparable clean forecasting accuracy. Additional evaluations on target-pattern generality, anomaly-detector stealthiness, architecturetransfer robustness, component ablations, and parameter sensitivity further support the effectiveness and robustness of the proposed frequency-aware design.
Deep learning (DL)-based multi-class intrusion detection systems (IDSs) in smart grids remain highly susceptible to stealthy backdoor attacks while the relevant defense research is still in its infancy. The existing activation clustering algorithm-based backdoor defense methods have successfully applied in the field of computer vision, but their design depends heavily on the experts’ experience. To overcome these limitations, this work originally proposes an evolutionary activation clustering combination-based backdoor defense method for Transformer-based multi-class IDSs in smart grids, called EvoACC-BDMIDS. It is the first time to present a BadNets-based backdoor attack sample generation mechanism for analyzing backdoor vulnerabilities in DL-based multi-class IDSs for smart grids, while strictly adhering to inherent the operational constraints of smart grids. Then, the automatic design of an optimal activation clustering combination strategy deployed for backdoor defense is formulated as a constrained single-objective optimization problem, where both integral variables, i.e., combination length and the type of clustering algorithms, and continuous variables, i.e., a combination of weight coefficients for the different combined clustering algorithms are considered as the decision variables. The designed objective function is to minimize backdoor attack success rate (BASR), constrained by maintaining the difference of detection performance on benign samples with and without backdoor defense below a defined threshold. To efficiently represent and evolve diverse combination strategies from 12 activation clustering algorithms, we employ a mixed variable-length encoding scheme and develop the tailored evolutionary operations under the framework of genetic algorithm. Experimental results on Power System Attack Datasets with 37 classes demonstrate the proposed EvoACC-BD-MIDS maintains 91.4% average accuracy and reduces BASR from 98.89% to 0.00%, and it significantly outperforms other single activation clustering-based backdoor defenses. The code is available at: https://github.com/SomeonesGitee/EvoACC-BD-MIDS.git.
Message-locked encryption (MLE) enables the cloud server to perform deduplication over encrypted data that users outsource. To resist brute-force dictionary attacks against data, a feasible solution is server-aided MLE, which employs dedicated key server(s) to aid users in generating encryption keys (i.e., MLE keys). Existing server-aided MLE schemes also consider portability, a practical property that allows users to access outsourced data from multiple devices using passwords. In this paper, we analyze existing server-aided MLE schemes and reveal that they are confronted with subverted implementation, where algorithms in server-aided MLE may be maliciously implemented to stealthily leak users’ secrets, including data, MLE keys, and passwords. In addition, these schemes fail to provide file sharing among users. To solve these issues, we propose a subversion-resistant portable MLE scheme with file sharing, dubbed STEAD. In STEAD, we conceive a hardened password-derived randomizable public-key encryption mechanism, where cryptographic reverse firewalls (CRFs) are deployed to rerandomize ciphertexts of MLE keys produced with users’ hardened password-derived public keys, thereby resisting subverted implementation while maintaining portability. CRFs also modify users’ incoming and outgoing messages during generation of MLE keys and hardened passwords to prevent secret leakage. Moreover, a file sharing mechanism is designed in STEAD.We formally prove the security of STEAD and conduct comprehensive performance evaluations.
The rapid progress of AI-generated image (AIGI) synthesis poses substantial challenges to AIGI detection, especially under cross-generator and cross-domain distribution shifts. We propose CLIP-BCP, a dual-branch framework that combines the mid-layer CLS token of a frozen CLIP encoder with a bright channel prior (BCP). On two benchmarks, we show that mid-layer CLIP features provide stronger real/fake separability while balancing semantic context and fine-grained artifacts. CLIP-BCP freezes the pretrained CLIP encoder and uses the 13th-layer CLS token as its semantic backbone. In parallel, the BCP branch encodes multiscale bright channel response distributions through morphological dilation and convolutional encoding. Feature-level fusion combines these distributions with the mid-layer CLIP representation. Experiments on AIGCDetect-Benchmark and GenImage show that CLIP-BCP achieves mean accuracy (mAcc) values of 94.28% and 98.29%, outperforming the previous best methods by 1.51 and 2.49 percentage points, respectively. Strict cross-dataset evaluation, postprocessing robustness tests, and ablations of the prior and fusion strategy further support the complementarity of mid-layer CLIP features and the BCP under the evaluated settings. The source code is publicly available at https://github.com/LeiChen0407/CLIP-BCP.
Cryptosystems based on Learning with Errors (LWE) are prominent candidates for post-quantum cryptography, but estimating their side-channel security remains challenging. Existing methods provide only loose bounds in most cases. In this paper, we develop a theoretical framework for side-channel analysis of LWE-based cryptosystems via the replica method, which provides a tractable asymptotic estimate of the security metrics for attacks formulated as distribution hints under commonly adopted statistical-physics assumptions. As a representative case, we analytically derive the security metrics for Kyber, a LWE-based key-encapsulation mechanism, under a typical inequality-hint-formed side-channel attack. Based on the theoretical results, we numerically determine the minimum number of attacks required to recover the secret key. We also extend the analysis to practical attack settings and broader hint models, demonstrating the generality and practical adaptability of the proposed framework. The experimental results are in close agreement with the asymptotic predictions and show that the proposed framework provides more accurate estimates than existing methods.
