The proliferation of highly realistic forged images, driven by advances in generative models such as GANs and diffusion models, poses serious challenges to information security and digital forensics. While existing detection methods leveraging spatial-domain features have shown progress, they often struggle to generalize to ever-evolving forgery techniques–particularly AI-generated content. To address this limitation, we propose a novel Multi-scale Fourier Enhancement Network that integrates frequency-domain analysis with multi-scale feature learning. Our approach incorporates a lightweight Fourier Enhancement Module into a ConvNeXt backbone, enabling the capture of frequency anomalies introduced by forgery operations. Combined with a multi-scale fusion mechanism and a hybrid loss function, the model achieves robust feature learning from local textures to global semantics. Extensive experiments on multiple benchmarks demonstrate that our method achieves superior accuracy and robustness while maintaining computational efficiency, and shows promising cross-generator generalization performance on unseen generative sources.
Abstract Embedding-cost design is central to distortion-minimization steganography, yet most GAN-based cost-learning methods are restricted to ternary modifications ( $$\{0,\pm 1\}$$ { 0 , ± 1 } ), which can become limiting at high payloads. We propose Curriculum-based Pentary Embedding GAN (CuPE-GAN) for learning pentary ( $$\{0,\pm 1,\pm 2\}$$ { 0 , ± 1 , ± 2 } ) embedding costs. CuPE-GAN uses a two-layer ternary training surrogate: Stage 1 learns the first-layer $$\pm 1$$ ± 1 allocation, whereas Stage 2 activates the second layer with forced learning-rate decay; a directional-consistency loss discourages canceling layerwise changes; and an Improved Decoupled Ternary Approximation (IDTA) represents the two modification directions with separate one-sided soft indicators and exact zero-probability consistency at finite $$\lambda$$ λ . The learned layer probabilities are converted to a normalized five-symbol distribution and then to pentary costs, while actual message embedding is performed by a standard pentary STC encoder. Trained on CelebA-HQ and ALASKA1 and evaluated on BOSSBase and BOWS2, CuPE-GAN achieves higher average detection error rates than Steg-GMAN, WOW, and S-UNIWARD at medium and high payloads. Relative to Steg-GMAN, the average gains at 0.4 and 0.8 bpp are 6.16 and 5.23 percentage points on BOSSBase, 5.05 and 7.92 percentage points on BOWS2.
Abstract Simeck demonstrates reliable resistance against conventional cryptanalytic attacks and offers sound security properties, making it a noteworthy design in the field of lightweight cryptography. Nevertheless, ongoing progress in quantum computing poses a potential threat to the security of Simeck, underscoring the need for a rigorous analysis of its quantum resilience. In this paper, we analyze quantum distinguishers and key-recovery attacks on Simeck in the Q2 model. Leveraging the connection between truncated differentials and periodic functions, we revisit Simon-type qCPA distinguishers for Simeck-32/48/64 and further extend them to the qCCA setting. The qCPA distinguishers cover 8/9/11 rounds of Simeck-32/48/64, matching the round numbers reported by Lu et al. (2026), whereas our qCCA distinguishers cover 10/11/13 rounds. Based on these distinguishers, we investigate their application to Grover-meets-Simon key-recovery attacks on reduced-round Simeck. Furthermore, we mount qCPA key-recovery attacks on 11/12/14-round Simeck-32/48/64 and extend them to qCCA attacks on 13/14/16-round Simeck-32/48/64. Finally, we provide empirical estimates of the quantum resources required for the proposed key-recovery attacks. For Simeck-32/48/64, the empirical quantum query complexities are $$2^{28.09}/2^{40.32}/2^{52.59}$$ 2 28.09 / 2 40.32 / 2 52.59 in the qCPA setting and $$2^{26.81}/2^{39}/2^{51.17}$$ 2 26.81 / 2 39 / 2 51.17 in the qCCA setting, respectively, under the stated decision-test assumptions and empirical false-positive estimates.
Abstract Multi-stage cyber attacks spanning extended periods pose significant challenges for security investigation, as analysts must reconstruct the complete attack sequence from voluminous system audit logs. Existing provenance graph-based investigation methods either rely on predefined attack knowledge to map activities to known stages, or analyze the graph from alert events using a single set of analysis criteria that fail to adapt to the behaviorally distinct phases of an attack. Consequently, attack activities that are temporally or causally distant from the triggering alert are often inadequately recovered. We observe that different attack stages produce distinct distributions of temporal subgraph patterns in provenance graphs, providing a basis for automated stage partitioning without prior attack knowledge. Based on this observation, we propose MotifTrace , a two-phase framework that automates multi-stage attack investigation on provenance graphs without predefined attack knowledge or labeled data. MotifEmbed learns edge-level representations that preserve stage-specific temporal subgraph structures and clusters them into behaviorally coherent stage partitions that support downstream reconstruction. PathRank then identifies critical attack paths within each stage through topology-aware causal path scoring, and merges them across stages through shared entities to reconstruct the complete attack scenario. Evaluations on six attack campaigns of varying complexity, including DARPA Transparent Computing benchmarks, demonstrate that MotifTrace achieves an average Macro- $$F_1$$ F 1 of 0.88 in multi-stage reconstruction, outperforming state-of-the-art methods by over 21%. By eliminating the need for manual stage annotation and predefined attack signatures, MotifTrace automates a labor-intensive phase of the security investigation workflow, reducing analyst effort from provenance graph construction to structured attack summaries.
Abstract Retrieval-augmented generation (RAG) substantially improves large language models (LLMs) on knowledge-intensive tasks by incorporating external knowledge, yet open knowledge sources introduce prominent security risks. The SafeRAG benchmark identifies four novel textual evidence manipulation attacks, namely silver noise, inter-context conflict, soft advertisement, and White Denial-of-Service (White DoS), which bypass conventional defense components and threaten both the security and availability of RAG systems. Existing single-stage defenses fail to capture the implicit malicious characteristics of these attacks, while multi-stage frameworks are limited by fixed parameters, weak conflict resolution, and insufficient detection of implicit toxicity and availability threats. To address these limitations, we propose MIRAGE, a two-stage adaptive defense framework combining retrieval-side purification and generation-side resolution. The first stage applies attack-characteristic-aware dynamic K-means clustering and dynamic ROUGE-L threshold filtering to eliminate malicious documents. The second stage employs attack-adaptive knowledge integration rules alongside an enhanced self-evaluation mechanism that jointly verifies factual accuracy, implicit toxicity, and availability. Under the static, non-adaptive attack setting defined by SafeRAG, experiments on DeepSeek-v3.2 show that MIRAGE achieves average improvements of 20.9% in Attack Failure Rate (AFR), 11.5% in Retrieval Accuracy (RA), and 14.1% in F1(avg) over baselines across all four attack scenarios, providing effective technical support for the stable deployment of RAG systems in complex adversarial environments.
