Our study systematically evaluates the robustness of fine-tuned pre-trained models across different datasets and attack types, aiming to provide insights for researchers in selecting pre-trained models for image classification tasks while prioritizing robustness against potential attacks. Our work explores the influence of fine-tuning pre-trained models with different datasets and attack types on their robustness and overview of the appropriate pre-trained model for image classification based on factors such as computation resources, memory usage, and accuracy on clean data. We fine-tune the models with weights. Evaluate the models on data before any attack and after different attack cases. The study utilizes publicly available image datasets such as MNIST, CIFAR10, CIFAR100, and SVHN to assess the robustness of pre-trained models. The results indicate that the robustness of fine-tuned models varies across different data distributions and sample sizes, emphasizing the need for tailored selection based on specific task requirements. The findings also highlight the varying performance of ResNet50, DenseNet, and MobileNet in defending against attacks on different datasets. Overall, this research provides valuable insights into the robustness of fine-tuned pre-trained models and their suitability for real-world applications in image classification.
Context: Non-determinism in deep learning algorithm design and implementation leads to performance variation, meaning model performance is not a single value, but rather a distribution. These model performance distributions are underexplored despite their impact on robustness. Objective: We investigate the robustness of deep learning performance to sources of non-determinism, specifically focusing on how performance distributions differ across various architectures and tasks. Method: We conducted 186 experiments on image classification and time series forecasting architectures. Each experiment was run 100 times with different random seeds to generate performance distributions, resulting in 18,600 runs. Robustness was quantified using metrics for spread, symmetry, and tail risk. Results: Performance distributions are frequently non-Gaussian, particularly in time series forecasting. Model size does not systematically affect robustness -- larger image classification models show fewer outliers but not lower spread, while smaller time series models show lower spread but more extreme underperformers. Training duration does not scale linearly; early stopping effectively balances performance and robustness. Mean performance does not predict robustness - time series forecasting shows moderate correlation while image classification shows none. Time series models produce nearly three times more underperforming outliers than image classification models, indicating substantially higher tail risk. Conclusions: Tail risk poses serious concerns for Trustworthy AI in high-stakes applications. Models performing well on average may exhibit long tails and extreme outliers revealed only through distributional analysis. Mean performance alone should not guide model selection; assessment of spread, symmetry, and tail risk is essential for reliable model assessment where consistent performance is critical.
Formula-Driven Supervised Learning (FDSL) enables vision pre-training without real images, but existing methods cast pre-training as classification on discretized labels, discarding continuity in the generative process. We propose Formula-Driven Supervised Metric Learning (FDSML), which derives relational supervision from image-generation parameters beyond discrete class labels. FDSML combines FineFractalDB, our refined fractal dataset with higher intra-class variation and more stable structures, with a Parameter-Aware Triplet (PA-Triplet) Loss that assigns positives and negatives according to distances in the underlying Iterated Function System parameter space. This objective trains embeddings to respect parameter-derived relative similarity orderings. Experiments show that FDSML improves transferability to image retrieval while remaining competitive on downstream classification and improving robustness under several corruption settings. On ResNet-50, Recall@1 improves from 65.6% to 74.0% on Stanford Online Products and from 56.2% to 78.4% on Stanford Cars over FDSL pre-training. These results suggest that parameter-derived relational supervision provides a stronger synthetic pre-training signal than label-based classification for transfer beyond recognition.
Few-Shot Learning (FSL) addresses classification scenarios in which only a small number of labeled examples are available for each class. This paper presents the Sum12 Multi-Scale branching encoder, a lightweight convolutional architecture designed to improve the embedding stage of Prototypical Networks without substantially increasing computational cost. The encoder uses seven parallel convolutional branches whose kernel-size combinations sum to 12, allowing local and broader spatial structures to be represented within a unified Multi-Scale feature space. The proposed architecture was evaluated on Omniglot, CIFAR- FS, FC100, and miniImageNet under controlled few-shot learning protocols and compared with the original Prototypical Network encoder and Inception-inspired branch variants. Additional ablation studies have investigated fusion strategies, dilated convolutions, branch compositions, and receptive field budget. Beyond few-shot learning, the encoder was also tested on the MedMNIST 2D benchmark to assess its behavior in conventional supervised medical image classification. Across these experiments, Sum12 generally improved the low-shot classification performance with respect to the basic ProtoNet encoder while preserving a compact model size. These properties suggest that the proposed encoder may be suitable for low-data visual recognition and resource-constrained intelligent imaging applications. However, complete end-to- end validation on the target embedded platform, including resolution of the observed implementation discrepancies, remains future work.
Multi-source remote sensing image classification benefits from complementary spectral and structural information. However, identifying discriminative cross-modal features remains challenging. This letter proposes a Cross-Layer Refinement and Frequency-aware Fusion Network (CRFNet) within a lightweight dual-branch architecture. Specifically, a Cross-Layer Refinement Attention (CLRA) module propagates selected attention relationships across successive layers and employs top relation selection to progressively strengthen discriminative spatial dependencies while suppressing irrelevant responses. Furthermore, a Frequency-aware Complementary Fusion (FCF) module combines frequency-wise interactions with cross-modal channel attention in the Fourier domain. The resulting representations are then converted into spatial-domain gates to selectively modulate complementary features. Experiments on the Houston 2013 and Augsburg datasets demonstrate that the proposed CRFNet consistently outperforms several state-of-the-art methods. The codes are publicly available at https://github.com/oucailab/CRFNet
Persistent homology (PH) is a mathematically principled multi-scale topological descriptor of shape and connectivity in data. It is widely applied in image classification, segmentation, and shape analysis across medical and aerial imaging, providing structural information that pixel-wise deep learning may overlook. Standard cubical PH decomposes an image into cubical cells (vertices, edges, and squares) and reduces them sequentially in filtration order. This sequential reduction has $\Theta(n)$ parallel depth for an $n$-pixel image, limiting within-image GPU parallelism and constraining its use in topology-aware deep learning. We introduce CuPH, a GPU-based parallel algorithm for computing cubical PH on 2D images that operates directly on pixels rather than cubical cells. CuPH buckets pixels by filtration values, activates them with a level-synchronous sweep, and tracks connected components with a lock-free union-find method. We prove that CuPH produces persistence diagrams identical to standard cubical PH on integer inputs, with a stability bound for floating-point inputs. CuPH reduces parallel depth from $\Theta(n)$ to $O(L\log n)$ for an image with $L$ intensity levels. We implement the level sweep as a single CUDA kernel, eliminating cross-level kernel-launch overhead. Empirically, CuPH is $18.6$--$99.6\times$ faster than the existing GPU baseline (CMP). As a drop-in PH backend, CuPH accelerates topology-aware classification ($4.2$--$16.7\times$) and segmentation ($8.3$--$44\times$) while preserving accuracy. It also enables computing full-image PH losses on high-resolution images, improving segmentation Dice by $2.2$--$4.6$ points over patch-local PH baselines and accelerating training by $25$--$31\times$.
The detection of previously unidentified Child Sexual Abuse Material (CSAM) remains a significant challenge for law enforcement authorities, online service providers, and child protection organizations, as traditional hash-matching technologies are primarily designed to identify previously verified material. This paper reviews and classifies existing AI-based technologies that could complement hash-based approaches and support the detection, analysis, and prioritization of potentially harmful content. The reviewed approaches are organized into three categories: real-time image and video analysis, image and video classification and tagging, and AI-based behavioral monitoring. A comparative analysis examines the capabilities, limitations, and potential integration of general-purpose computer vision platforms, content moderation services, and specialized CSAM detection tools within existing detection workflows. An experimental evaluation was conducted using Google Cloud Vision AI and Sightengine to investigate sexual-content classification and child-adult age estimation. The results highlight differences in classification taxonomies and outputs across providers, as well as significant limitations in age classification, particularly for adolescents. These findings highlight the importance of using AI-based assessments as complementary tools for analysis and prioritization, rather than as definitive determination of CSAM. Finally, the paper examines the relevant European Union regulatory framework, with particular attention to the EU AI Act, the General Data Protection Regulation (GDPR), fundamental rights, transparency, accountability, and human oversight. The legal analysis highlights how regulatory obligations and safeguards may differ depending on the deployment context, potentially creating inconsistencies across interconnected detection and reporting workflows.
Task-oriented semantic communication transmits task-relevant features to support receiver inference under limited communication resources and adverse channel conditions. For image classification, the usefulness of a feature depends on the other transmitted features and the distortion introduced by the wireless channel. Effective feature prioritization should therefore account for classification performance after transmission of the selected features. This paper proposes a Task-Oriented Feature Selection (\textit{TOFS}) framework and method, which evaluates feature replacements using receiver classification loss. Loss differences confirmed under independent transmission samples provide pairwise ranking supervision for a selector network. The trained selector predicts priorities in a single forward pass. CIFAR-10 simulations demonstrate improved classification accuracy over an importance-based feature selection baseline at the same transmitted feature count and average transmit power. The proposed \textit{TOFS} framework shows 84\% accuracy performance under Rayleigh fading and Gaussian noise, compared with the 68\% performance of a SOTA method.
Vision Transformers (ViTs) are bottlenecked by the length of their input token sequence. Standard patchification fixes this length with a single patch size applied across the image, regardless of regional content. Prior work on adaptive patch sizing reduces token count by assigning coarser patches to visually simple regions, but relies on noisy pixel-level heuristics such as local entropy to make patchification decisions. We argue that this compute allocation should be grounded in the model's interpretation of task relevant information, not in raw pixel statistics. We introduce Split Policy Learned Image Tokenization which learns to allocate token resolution based on what the model finds meaningful, producing shorter token sequences that are better matched to task-relevant content. SPLIT matches the accuracy of full-resolution ViT classification baselines while using 30% fewer tokens. The same gains hold on dense prediction: SPLIT achieves competitive accuracy on COCO and ADE20K, demonstrating that semantically aligned patchification generalises well beyond classification.
Background/Objectives: CNNs perform well in retinal disease classification from spectral-domain optical coherence tomography (SD-OCT), but whether EfficientNet's natural-image scaling hierarchy transfers to medical-image transfer learning at fixed resolution remains unclear. This study evaluates EfficientNet-B0-B7 for four-class SD-OCT classification and accuracy-efficiency trade-offs. Methods: Eight ImageNet-pretrained variants (4.05-64.10M instantiated backbone parameters) were evaluated at 224x224 on the Kermany dataset. Grayscale images were replicated to RGB and processed using model-embedded preprocessing and augmentation. Training comprised frozen-backbone head training followed by partial fine-tuning. Metrics covered classification, discrimination, efficiency, and visual analyses. Results: Test accuracy varied non-monotonically from 95.93% (B0) to 96.95% (B1); B7 achieved 96.73%. B1 had the highest accuracy and Drusen F1 (0.8780), while B7 had Drusen F1 of 0.8720 and the highest Drusen recall (0.9177). Drusen had the lowest F1 across all variants and contributed 82.2% of the summed squared deviation in class-specific F1; class ranking remained Normal > CNV > DME > Drusen. The OCT accuracy ranking differed from the reported ImageNet hierarchy, where B7 ranked first. The accuracy-to-parameter ratio fell from 19.82 to 1.48 percentage points per million total parameters. The final train-test gap had a weak, non-significant association with test accuracy (r = 0.26, p = 0.53). Conclusions: Larger EfficientNet variants did not consistently improve SD-OCT classification. B1 achieved the highest accuracy and Drusen F1, whereas B7 showed a secondary accuracy recovery at greater model size and CPU latency, indicating task and protocol dependent scaling effects. As cross-partition duplicates have not been independently audited, results are conditional on the supplied split.
Score-based multimodal music retrieval requires representations that are both discriminative within each modality and compatible across audio recordings, sheet music images, and symbolic scores. This paper investigates the effectiveness of parameter-efficient fine-tuning (PEFT) for adapting pretrained music encoders and examines whether modality-specific improvements translate into better multimodal representations. We compare full fine-tuning, LoRA, and QLoRA across nine pretrained audio, image, and symbolic score encoders on TriScore and the Musical Theme Dataset. Encoders are first evaluated and selected through supervised unimodal classification, using frozen-encoder linear probing as a reference. We then introduce a controlled 2x2 design crossing encoder adaptation with task-specific shared projection, comparing frozen and adapted representations before and after projection. LoRA or QLoRA is selected in 14 of 18 modality-target combinations, frequently matching or outperforming full fine-tuning while optimizing substantially fewer parameters. However, these unimodal gains do not transfer uniformly to multimodal classification. Adaptation and projection are complementary for composer classification on TriScore, largely redundant for TriScore instrument and the Musical Theme Dataset instrument and ensemble tasks, and partially interfering for composer classification on the Musical Theme Dataset. The final adapted-and-projected title representation on TriScore supports piece-level retrieval, reaching average Top-5 and Top-10 accuracies of 55.32 and 64.68, respectively. Cross-modal exclusion and neighborhood analyses further reveal strong modality-specific local structure in the projected spaces. Overall, PEFT provides an effective and computationally efficient alternative to full fine-tuning, but improved unimodal discrimination alone does not guarantee better multimodal organization.
We present BDNet, a hierarchical image-classification backbone built around fast image features and slow statistical references. Its baseline deformation operator (BDO) transforms the deviation of a feature from a stage-specific reference through a low-rank projection and a learned channel scale. Rather than computing pairwise token similarity, the branch operates directly on baseline-relative residuals with complexity linear in the token count. Each of four reference buffers accumulates detached stage statistics during training and remains frozen at inference. A four-stage pyramid with depths $(2,2,6,2)$ and widths $(96,192,384,768)$ yields a 20.70M-parameter classifier at $224\times224$ resolution. We organize an ImageNet-1K study around clean classification, eight controlled ablations, and Gaussian-noise, contrast, and brightness perturbations. The central hypothesis is that a slow reference can improve the robustness of a compact representation while retaining a simple tensor implementation. We distinguish this training-time statistical mechanism from input-dependent attention and make its linear inference behavior explicit.
Diffusion-based data augmentation has emerged as a powerful strategy for addressing data-scarce image classification. Recent pipelines typically adapt a pretrained text-to-image diffusion model to a limited number of class examples and subsequently sample synthetic images using diverse textual prompts. However, in extreme data-scarce regimes such as long-tailed single-sample settings, this approach suffers from a dual failure mode: the generated samples often lack discriminative foreground details crucial for classifier training, and the adapted class prior can bias the early denoising process, causing output images to exhibit similar patterns despite varying text prompts. To overcome these limitations, we propose a sampling-side framework of two complementary components that simultaneously enhance fine details and diversity. First, Mask-Guided Reference (MGR) Attention updates only foreground value tokens with reference features via zero-shot segmentation, limiting reference influence to the target object and mitigating excessive global feature injection. Second, Repulsive Noise Optimization (RNO) encourages initial noise trajectories to diverge from dominant generative modes of the adapted model, enabling target prompts to express broader contextual variation. Extensive evaluations show consistent improvements, with the largest gains in long-tailed single-sample settings of up to 2.50 and 2.94 points on CUB-LT IF100, and further robustness across few-shot and background-shift evaluations.
Accurate localization of abnormalities in three-dimensional (3D) CT scans is essential for linking automated predictions with clinically relevant pathological findings. We propose Uni-Loc, a vision-language framework that unifies pathology classification and 3D localization using scan-level supervision. Uni-Loc derives both predictions and localization maps from spatial similarities between pathology-conditioned queries and image features. A token-to-voxel upsampling module enables high-resolution localization, while gradient-guided thresholding identifies pathology regions without additional learned parameters. Evaluated on five localization tasks comprising 355 annotated CT scans, Uni-Loc achieves macro-averaged voxel-level AUROC/AUPR of 96.1\%/34.5\%, outperforming relevance propagation (80.7%/13.9%) and gradient saliency (88.7%/8.1%). Uni-Loc also achieves classification AUROC of 88.39% on CT-RATE and 78.03% on RAD-ChestCT, improving over a matched baseline. These results demonstrate effective 3D pathology localization without requiring pathology-specific voxel-level supervision.
Real-time disaster recognition from aerial imagery requires deep learning models that operate under strict memory, compute, and latency constraints. While recent approaches achieve high accuracy, their computational cost limits deployment on resource-constrained UAV platforms. This work introduces picoEmergencyNet, an ultra-lightweight convolutional neural network designed for aerial image disaster classification. The proposed architecture combines depthwise dilated convolutions with efficient channel attention and channel shuffling to enable multi-scale feature extraction at minimal computational cost. picoEmergencyNet has 23k parameters and 7M FLOPs, requiring only 0.1 MB of model storage. Evaluated on the AIDER and AIDERv2 disaster datasets, it achieves up to 96.1% accuracy, matching or exceeding state-of-the-art lightweight models that are orders of magnitude larger. Deployment experiments on the NVIDIA Jetson AGX Orin show that, without hardware-specific optimizations, picoEmergencyNet achieves 258 FPS on GPU and 46 FPS on CPU with a 7x lower activation memory footprint compared to competing models. These results demonstrate that careful architectural specialization, rather than model scaling, enables highly efficient and accurate TinyML-ready disaster recognition for real-time UAV-based emergency monitoring.
Modern image classification models largely rely on local characteristics and feature extraction for their predictions. This focused approach works well in the spatial domain, but reportedly falls short in the frequency domain, even though both representations contain the same information. This study demonstrates that the significant reduction in accuracy in the frequency domain stems from the lack of local features. Locally swapping data in the frequency domain significantly decreases its inverse spatial domain accuracy, even below that of the frequency domain; however, the frequency domain accuracy remains relatively constant. Moreover, the use of datasets containing global features in the spatial domain yields higher accuracy in the frequency domain. Using these results, a hybrid model is constructed, combining one model trained only in the frequency domain and another trained only in the spatial domain to achieve higher classification accuracy. The proposed visual geometry group (VGG) and vision transformer (ViT) architectures are trained on the ImageNet-100, WikiArt, DTD, FMD, PSVRT, and KTH-TIPS2-b datasets and evaluated.
Kernel machines are an important class of models in machine learning. However, the question of how robust modern kernel machines are to adversarial examples has not been studied comprehensively. In this paper, we provide a thorough empirical investigation of the adversarial robustness of kernel-based models for both binary and multiclass classification tasks, via white-box and black-box evasion attacks. We evaluate the native robustness of kernel machines, random feature classifiers, and the relatively newer Recursive Feature Machines. In addition, we provide an algorithm for approximate adversarial training of kernel machines. Our experiments on classification tasks on image and tabular datasets strongly indicate that kernel machines are natively more robust than neural networks, i.e., their performance degrades more gracefully against adversarial perturbations.
Remote sensing scene classification (RSSC) is traditionally formulated as a vision-centric problem, where inference relies on rich visual representations extracted from images. In this work, we revisit this assumption by introducing text-only RSSC, a new problem setting that investigates whether remote sensing scenes can be recognized using only textual semantic representations. To enable this study, we propose SeProRS, which reformulates RSSC as a text-only classification task through semantic proxy generation. Specifically, a vision-language model (VLM) is first employed to convert remote sensing images into textual descriptions, termed semantic proxies, which are then encoded by a text embedding model for scene classification. This decoupled pipeline provides a unified and reproducible benchmark for evaluating text-only scene understanding. Furthermore, we investigate representative prompting strategies and proxy generators, and analyze their impact on classification performance, text embedding behavior, cross-generator transferability, and efficiency. Extensive experiments on three well-known RSSC datasets show that semantic proxies can preserve substantial discriminative information, while also revealing inherent limitations of purely text-based scene understanding. This paper does not aim to replace image-based RSSC, but instead establishes a new research direction by quantifying the feasibility and boundaries of semantic proxy-based scene recognition. The proposed benchmark and analysis provide a foundation for future research on efficient and modality-reduced remote sensing scene understanding. The source code and datasets will be accessible at https://github.com/sstary/SSRS.
DINO (self-DIstillation with NO labels) has become a stan- dard for self-supervised pre-training, producing features that generalise across datasets and downstream tasks. In medical imaging it has been adopted mainly for radiology (X-ray, CT and MRI), where even natural- image DINO features transfer well. In the literature, these features have been reported to transfer worse to functional imaging such as PET. Yet PET-specific DINO models have barely been introduced, and even less validated across multiple tasks and datasets. We present PET-DINOv2-2D, the first DINOv2 model pre-trained on 2D PET, and benchmark it against open-source encoders: the origi- nal DINOv2 (pre-trained on the LVD-142M natural-image dataset) and OmniRad (pre-trained on CT, MRI and ultrasound). Using the same lightweight heads for every backbone, and despite far less pre-training data, PET-DINOv2-2D matches or outperforms both on three down- stream PET tasks: normal/abnormal classification, lesion segmentation and recurrence-free survival prognosis. Its advantage is largest when few labels are available, suggesting that PET pre-training gives more transferable, label-efficient features. Code is made available at https://github.com/nuclivision/pet-dinov2-2d/.
Recent advances in generative models have made synthetic images increasingly photorealistic, posing growing challenges to reliable and generalizable detection of AI-generated images. Existing methods typically rely on holistic image-level classification, which may overlook local artifacts, especially in high-resolution images. To address this limitation, we formulate this task as a falsification-driven multiple instance learning problem, where a single fake patch can falsify the whole image, while a real image requires all patches to be authentic. Based on this formulation, we propose \textit{Learning to Falsify (LTF)}, a hierarchical multiple instance learning framework that models an image as multiple bags of instances across different spatial levels and aggregates their predictions under the falsification-inspired principle. We further introduce Cross-Shuffle during training to encourage the model to learn local artifact evidence rather than image-level shortcuts. Furthermore, we propose the \textit{Falsification-driven Test-time Aggregation Strategy (FATAS)}, a plug-and-play module that extends the same falsification-inspired principle to existing detectors by aggregating patch-level predictions. Extensive experiments demonstrate that LTF achieves strong generalization and robustness across multiple benchmarks, while FATAS substantially improves representative detectors in high-resolution scenarios without retraining. Code is available at https://github.com/HuiGuanLab/LTF.git