Exploring Adversarial Watermarking in Transformer-Based Models: Transferability and Robustness Against Defense Mechanism for Medical Images

A new watermarking paradigm, RDSplat, has been introduced to enhance the robustness of digital watermarking against diffusion-based editing in 3D Gaussian Splatting (3DGS). This method embeds watermarks into components that are preserved during diffusion editing, addressing vulnerabilities in existing techniques.

Recent research has identified an 'Inductive Bottleneck' in Vision Transformers (ViTs), where these models exhibit a U-shaped entropy profile, compressing information in middle layers before expanding it for final classification. This phenomenon is linked to the semantic abstraction required by specific tasks and is not merely an architectural flaw but a data-dependent adaptation observed across various datasets such as UC Merced, Tiny ImageNet, and CIFAR-100.

A new framework called Cross-modal Explainable Framework for Melanoma (CEFM) has been introduced, utilizing contrastive learning to enhance interpretability in melanoma diagnosis by aligning clinical criteria with visual features through Vision Transformer embeddings.

A new study introduces the Asynchronous Bioplausible Neuron (ABN), a dynamic spike firing mechanism designed to enhance Spiking Neural Networks (SNNs) for computer vision applications. This innovation addresses the challenge of maintaining homeostasis in neural networks by auto-adjusting to variations in input signals, leading to improved image classification and segmentation performance.

A recent study evaluates the sensitivity of Bidirectional Long Short-Term Memory (BiLSTM) forecasting models to input sequence length and noise, highlighting their effectiveness in time-series forecasting across various domains, including environmental monitoring and the Internet of Things (IoT).

The Universal Weight Subspace Hypothesis reveals that deep neural networks, including Mistral-7B, Vision Transformers, and LLaMA-8B, converge to similar low-dimensional parametric subspaces across various tasks and domains. This study provides empirical evidence from over 1100 models, indicating that neural networks exploit shared spectral subspaces regardless of their initialization or the specific task they are trained on.

The subspace-constrained Tyler's estimator (STE) has been analyzed for its effectiveness in recovering low-dimensional subspaces from datasets affected by outliers. This method has shown competitiveness in key computer vision tasks, particularly under conditions where the inlier fraction is low, which complicates robust subspace recovery. The study establishes that with proper initialization, STE can efficiently recover the underlying subspace even in challenging scenarios.

The introduction of AutoNeural marks a significant advancement in the design of Vision-Language Models (VLMs) specifically optimized for Neural Processing Units (NPUs). This architecture addresses the inefficiencies of existing VLMs on edge AI hardware by utilizing a MobileNetV5-style backbone and integrating State-Space Model principles, enabling stable integer-only inference.