Vision Transformers (ViTs) are vulnerable to adversarial patch attacks, posing serious challenges to their deployment in security-critical applications. Existing defense strategies improve robustness but at the cost of significant computational overhead, limiting their practicality. To address this, we propose STEM, a lightweight and parallelizable token pruning defense mechanism that enhances robustness while reducing inference cost. STEM integrates two components: Block-wise Selective Fusion (BSF), which fuses redundant tokens and flags suspicious ones, and Identify-Prune (IP), which identifies and prunes adversarial tokens based on multi-head attention statistics.To support STEM efficiently in hardware, we further design STEMA, a scalable accelerator featuring a dedicated Security Core operating in parallel with the ViT backbone. STEMA includes specialized engines for token fusion, top-k selection, and metadata management, occupying only 5% of chip area.Experimental results show that STEM achieves state-of-the-art (SOTA) defense performance and demonstrates excellent efficiency. Specifically, STEM improves robustness by up to 4.7× compared to token pruning methods and achieves 46.8×–105.8× runtime improvements over defense baselines. STEMA further delivers up to 71× speedups and 397× energy efficiency improvements, demonstrating its suitability for secure and efficient ViT deployment, especially in edge environments.

Multi-view learning methods integrate multiple views to improve performance. Existing methods typically rely on highly reliable views, while neglecting the potential exploitation to not entirely reliable ones. In this paper, we propose a novel Trustworthy Localized Corrections-guided Mutual Learning(TLCML) method to address this limitation in a fine-grained manner. Our method maximizes the exploitation of complementarity across views through mutual learning guided by trustworthy localized corrections. To achieve this, we first employ evidential neural networks to capture view-specific opinions, including evidence and uncertainty, where evidence denotes the support for categories. Next, a coupled dual-branch strategy is applied at the evidential element level to explore both intra-and inter-opinions correlations. This strategy facilitates mutual learning among trustworthy localized decisions to mitigate noise interference. Extensive experiments demonstrate that TLCML achieves state-of-the-art performance on five benchmarks. The code is available at https://github.com/qiuranl/papercode.

Exploration in sparse reward environments is a long-standing challenge in reinforcement learning. While advanced methods often enhance exploration by introducing intrinsic rewards, they may lead agents to local optima in procedurally-generated environments with high uncertainty and diversity. Specifically, agents may over rely on certain intrinsic signals, neglecting long-term tasks tied to external rewards. To address this, we leverage multimodal causal reasoning to identify key causal variables relevant to external tasks and propose a Causal Reasoning-guided intrinsic Goal generation method (CRG). Our approach integrates multimodal data to abstract invariant causal relationships in dynamic environments, using a teacher network to evaluate and select intrinsic goals with the strongest causal impact on task completion. This enables agents to explore and learn efficiently, enhancing overall task performance. We evaluated our method in several MiniGrid environments of varying complexity. Experimental results show that CRG can leverage causal relationships to generate more effective internal goals, significantly accelerating task completion and improving exploration efficiency.

Vision Transformer (ViT) has achieved remarkable performance in image classification but remains susceptible to natural corruptions and adversarial attacks. These attacks fabricate either global image-wide or localized patch-based perturbations. Although ViT demonstrates inherent resilience to natural corruptions and global image-wide adversarial attacks, it is still vulnerable to patch-based attacks. Current defense methods for ViT against such attacks typically involve re-inference or layer-by-layer analysis, incurring substantial computational costs. To address this issue, we introduce ViTProbe, a defense mechanism that identifies and removes adversarial patches within a single layer of the model. It is observed that the relative length of attention score associated with an adversarial patch is extremely low. Leveraging this insight, ViTProbe detects adversarial patches by assessing the ratio between components of the attention score matrix and the corresponding components of the input matrix in the model’s specific layer. Potential adversarial patches are discarded before forwarding the chosen layer’s output, thereby eliminating the need for re-inference. Comprehensive evaluations against three patched-based attacks demonstrate that ViTProbe can recover the classification accuracy under adversarial patch attacks to 94.17% of the clean accuracy on average, achieving performance comparable to state-of-the-art defense while significantly reducing computational costs by up to 95.67%.

Multiplication plays a critical role in SRAM-based Computing-in-Memory (CIM) architectures. However, current SRAM-based CIMs face three major limitations. First, they do not fully exploit bit-level sparsity, resulting in unnecessary overhead in both latency and energy consumption. Second, the generation of numerous zero-dot products is superfluous. Third, the irregular organization of SRAM complicates the implementation. To address these issues, we propose Shift-CIM, a general-purpose approach that fully leverages bit-level sparsity within SRAM-based multiplications. Shift-CIM aligns the multipliers within the SRAM array, accumulating only the required dot products based on the non-zero bits of the multipliers. Shift-CIM achieves a regular SRAM organization by assembling two irregular SRAM arrays in a transposed manner. Our evaluations show that Shift-CIM is highly efficient, operating at a supply voltage of 0.9 V and a frequency of 833 MHz, while incurring only a 4.8% area overhead. Despite these modest requirements, Shift-CIM significantly accelerates multiplication operations, achieving up to 3.08× the performance improvement and a 60% reduction in energy consumption compared to state-of-the-art designs.