Shellcode injection allows attackers to execute arbitrary code inside a target process without writing files to disk, and remains a common technique in modern malware and advanced persistent threats. Existing defenses cover only part of this threat landscape: Data Execution Prevention (DEP) prevents a process from executing its own data pages, and Arbitrary Code Guard (ACG) restricts dynamic code operations initiated by protected processes. However, neither mitigation prevents a remote process from acquiring executable memory in a target process through cross-process system calls. This paper presents XMemGuard, a kernel-mode detection framework that enforces the cross-process write-XOR-execute ( W â X ) principle: cross-process system calls must not create attacker-controlled executable memory in the target process. XMemGuard employs a pre-dispatch system-call interception mechanism based on Windows Event Tracing (ETW). This mechanism intercepts three categories of system calls responsible for cross-process acquisition of executable memory, removes the execution bit during the acquisition stage, and then uses an in-process monitoring component to capture the runtime outcomes of subsequent execution attempts. Compared to three existing detection methods, XMemGuard can detect all 33 manually constructed PoC variants, while the strongest baseline method can only detect 30. On in-the-wild samples, it achieves the highest accuracy, recall, and F1 score among the compared methods, with 99.79
Abstract Privacy-preserving machine learning (PPML) techniques enable collaborative machine learning tasks without exposing raw data of the participating parties. However, the efficient evaluation of non-polynomial functions in PPML remains a key challenge. In this work, we propose NASE (Neat and Accurate Secure Exponentiation), an MPC protocol tailored to secure radial basis function (RBF) kernel evaluation. By exploiting the fact that the exponent in the RBF kernel is always non-positive, NASE admits a lightweight construction that reduces computation overhead by up to 1.65 $$\times $$ × and communication overhead by up to 3.97 $$\times $$ × in the RBF kernel evaluation task on our datasets compared to SIRNN (IEEE S& P 2021), the prior SOTA secure exponentiation framework that has been used in prior work for secure RBF kernel evaluation. Taking NASE as the building block, we propose a privacy-preserving two-party RBF kernel SVM training protocol. Using the BFV scheme and MPC techniques, we introduce group-batch sampling for efficient sampling in ciphertext with a provable convergence guarantee, and propose a partial rotation method tailored to our scenario to optimize dot product computation. Experiments on our real-world datasets show that our protocol achieves model accuracy comparable to plaintext training, and the reproduced comparison with previous work shows reduced training runtime under the evaluated LAN and WAN settings in the semi-honest model.
Abstract Fuzzing stateful security protocols presents a significant technical challenge. The behavior of the target server depends not only on the content of individual messages but also on the evolution of its internal state across message sequences. This statefulness imposes severe limitations on traditional code-coverage-guided fuzzing techniques. As a result, effective fuzzing of stateful protocols requires state-aware strategies, which may originate from manual guidance, automated analysis, or a combination of both. Automated analysis methods typically assume that researchers lack prior knowledge of the protocol and rely on heuristic approaches to infer a protocol state machine during interactions, which is then used to guide the fuzzing process. However, in more common practical scenarios, researchers often have partial knowledge of the protocol specification or implementation, enabling them to manually construct parts of the state machine. When conflicts arise between automatically inferred states and expert-defined ones, current approaches lack effective mechanisms to integrate human prior knowledge into the state machine refinement process. To address the above challenges, this paper proposes AFLDL, a state-aware grey-box fuzzing approach based on deep learning. AFLDL leverages neural networks to establish mappings between memory snapshots and protocol states derived from prior knowledge. During fuzzing, it infers the current protocol state of the target system by analyzing runtime memory snapshots, and discovers new states through clustering analysis. These inferred states are used to progressively build a complete state machine representation that guides state-aware fuzzing. Experimental results demonstrate that AFLDL outperforms state-of-the-art fuzzing tools, achieving higher code coverage and triggering more crashes within the same time frame.
Abstract Prediction-based min-entropy estimators often assume that one target-bit setting is suitable for sources with different marginal and temporal structures. We present MATE, a target-bit-aware dual-view framework that combines temporal features with synchronized Gramian Angular Field, Markov Transition Field, and Recurrence Plot representations. MATE profiles target-bit settings from 1 to 16 and evaluates whether the resulting operating points remain stable after compression. Experiments on 18 sources spanning distribution-based, generalized binary autoregressive (gbAR), and deterministic families show a clear source-dependent pattern: all four gbAR sources and all ten $$M^*$$ M ∗ -sequence sources achieve their lowest profiled errors at 1 bit, whereas the distribution-based sources prefer settings between 9 and 12 bits. The mean best-profiled error is 0.074395, compared with 0.647652 at the fixed 8-bit setting. Pruning, 2-bit quantization, and their combination preserve the empirically preferred target-bit setting for all 18 sources. In the shared-stream evaluation under the controlled fixed-8-bit protocol, Full MATE attains a macro mean absolute error of 1.904 bits per symbol, numerically close to 1.910 for NIST SP 800-90B and lower than 3.197 for TPA-LSTM, although none of the prespecified source-level comparisons is significant after Holm correction. Under strict budget matching, HP-UJ has a slightly lower numerical macro error than MA-Joint, but the paired difference is not significant after Holm adjustment. Deployment results show substantial storage reduction and lower incremental CPU memory, but no consistent CPU/GPU acceleration or GPU-memory reduction. Overall, the empirically preferred target-bit granularity varies across the evaluated sources, and the profiled operating points remain stable under compression, while budget-matched accuracy and runtime efficiency require separate evaluation.
Abstract Concept drift poses a major challenge for long-lived malware family classifiers. While drift-aware systems have been proposed, most remain opaque, flagging or adapting to drift but rarely explaining why it occurs. This study extends a triplet-autoencoder and DBSCAN-based drift detection framework with explainability and validation components designed to analyze feature- and behavior-level changes underlying drift decisions. Static Windows PE binaries are represented using EMBER features and embedded into a discriminative latent space, where family-specific DBSCAN clusters define decision regions and per-family distance thresholds. We apply Integrated Gradients (IG) to explain drift decisions at the feature level and introduce a counterfactual validation strategy to assess whether these explanations correspond to causally meaningful changes. We define the Drift Correction Rate (DCR) to quantify explanation consistency. IG is compared against the model-agnostic baselines LIME and SHAP. Separately, we conduct an independent behavioral analysis using capabilities extracted by CAPA and their mappings to MITRE ATT& CK techniques to examine whether samples flagged as drifted also exhibit observable behavioral differences. Experiments on a malware dataset spanning multiple Windows families show that the detector flags covariate drift in approximately 30% of post-training samples. IG-based explanations achieve near-perfect drift correction across families and consistently outperform SHAP and LIME. Behavioral similarity metrics based on the Jaccard index and Jensen–Shannon divergence further indicate that drifted samples diverge from the training distribution not only in feature space but also in extracted capabilities and MITRE ATT& CK techniques. Overall, the results show that the proposed framework provides empirically validated feature-level explanations of latent-space drift, while the independent behavioral analysis offers additional evidence that many drifted samples also differ in observable malware capabilities and MITRE ATT& CK techniques.
