Deep neural networks are vulnerable to adversarial perturbations, limiting deployment in safety-critical applications such as synthetic aperture radar (SAR) automatic target recognition (ATR). Randomized smoothing improves robustness by averaging predictions over noisy inputs, but isotropic noise often fails to preserve the semantic structure of SAR imagery. We propose semantic smoothing, a defense that replaces noised-based perturbations with structured randomized transformations generated by a novel view synthesis model. For SAR, we condition on acquisition geometry to synthesize multiple plausible radar views. Predictions across generated randomized views are aggregated to form a robust classifier. Experiments show that semantic smoothing improves robustness against standard attacks, such as FGSM and PGD, and SAR-specific attacks, such as OTSA and SMGAA, while also increasing clean classification accuracy. These results demonstrate that randomized smoothing via semantically preserving geometric transformations is a promising alternative to isotropic noise for adversarial defense in structured sensing domains.

Deep convolutional and transformer-based detectors achieve strong performance for SAR ship detection but are often computationally prohibitive for real-time or onboard deployment. Lightweight models offer improved efficiency yet struggle to capture the complex structural relationships inherent in SAR backscatter. Most existing SAR knowledge-distillation approaches rely on feature or logit matching, which enforces localized activation similarity while neglecting the geometric relationships among object representations. We propose a Structured Unified Relational knowledGE distillation framework for SAR Ship detection (SURGE) that transfers relational geometry from a powerful teacher detector to a compact student detector using a contrastive InfoNCE objective in a shared projection embedding space. To the best of our knowledge, this work presents the first transformer-based SAR ship detector knowledge distillation framework in SAR domain. The framework is architecture-agnostic in the sense that it provides a common region-level distillation interface for two-stage, one-stage and transformer-based detectors without modifying their deployed architectures. Experiments on the SSDD and HRSID benchmarks demonstrate that the proposed method yields substantial improvements for two-stage detectors, achieving up to 6.2 mAP and 8.0 AP75 gains over baseline student and even surpassing teacher performance.

Vision Transformers (ViTs) have emerged as powerful architectures in medical image analysis, excelling in tasks such as disease detection, segmentation, and classification. However, their reliance on large, attention-driven models makes them vulnerable to hardware-level attacks. In this paper, we propose a novel threat model referred to as Med-Hammer that combines the Rowhammer hardware fault injection with neural Trojan attacks to compromise the integrity of ViT-based medical imaging systems. Specifically, we demonstrate how malicious bit flips induced via Rowhammer can trigger implanted neural Trojans, leading to targeted misclassification or suppression of critical diagnoses (e.g., tumors or lesions) in medical scans. Through extensive experiments on benchmark medical imaging datasets such as ISIC, Braib Tumor, and MedMNIST, we show that such attacks can remain stealthy while achieving high attack success rates about 82.51% and 92.56% in MobileViT and SwinTransformer, respectively. We further investigate how architectural properties, such as model sparsity, attention weight distribution, and number of features of the layer, impact attack effectiveness. Our findings highlight a critical and underexplored intersection between hardware-level faults and deep learning security in healthcare applications, underscoring the urgent need for robust defenses spanning both model architectures and underlying hardware platforms.

Forget and Rewire (FaR) methodology has demonstrated strong resilience against Bit-Flip Attacks (BFAs) on Transformer-based models by obfuscating critical parameters through dynamic rewiring of linear layers. However, the application of FaR introduces non-negligible performance and memory overheads, primarily due to the runtime modification of activation pathways and the lack of hardware-level optimization. To overcome these limitations, we propose FaRAccel, a novel hardware accelerator architecture implemented on FPGA, specifically designed to offload and optimize FaR operations. FaRAccel integrates reconfigurable logic for dynamic activation rerouting, and lightweight storage of rewiring configurations, enabling low-latency inference with minimal energy overhead. We evaluate FaRAccel across a suite of Transformer models and demonstrate substantial reductions in FaR inference latency and improvement in energy efficiency, while maintaining the robustness gains of the original FaR methodology. To the best of our knowledge, this is the first hardware-accelerated defense against BFAs in Transformers, effectively bridging the gap between algorithmic resilience and efficient deployment on real-world AI platforms.

Deep learners tend to perform well when trained under the closed set assumption but struggle when deployed under open set conditions. This motivates the field of Open Set Recognition in which we seek to give deep learners the ability to recognize whether a data sample belongs to the known classes trained on or comes from the surrounding infinite world. Existing open set recognition methods typically rely upon a single function for the dual task of distinguishing between knowns and unknowns as well as making known class distinction. This dual process leaves performance on the table as the function is not specialized for either task. In this work, we introduce Cascading Unknown Detection with Known Classification (Cas-DC), where we instead learn specialized functions in a cascading fashion for both known/unknown detection and fine class classification amongst the world of knowns. Our experiments and analysis demonstrate that Cas-DC handily outperforms modern methods in open set recognition when compared using AUROC scores and correct classification rate at various true positive rates.

Even though convolutional neural networks (CNNs) can classify objects in images very accurately, it is well known that the attention of the network may not always be on the semantically important regions of the scene. It has been observed that networks often learn background textures, which are not relevant to the object of interest. In turn this makes the networks susceptible to variations and changes in the background which may negatively affect their performance.We propose a new three-step training procedure called split training to reduce this bias in CNNs for object recognition using Infrared imagery and Color (RGB) data. Our split training procedure has three steps. First, a baseline model is trained to recognize objects in images without background, and the activations produced by the higher layers are observed. Next, a second network is trained using Mean Square Error (MSE) loss to produce the same activations, but in response to the objects embedded in background. This forces the second network to ignore the background while focusing on the object of interest. Finally, with layers producing the activations frozen, the rest of the second network is trained using cross-entropy loss to classify the objects in images with background. Our training method outperforms the traditional training procedure in both a simple CNN architecture, as well as for deep CNNs like VGG and DenseNet, and learns to mimic human vision which focuses more on shape and structure than background with higher accuracy.

Neural network compression techniques almost always operate on pretrained filters. In this paper we propose a sparse training method for simultaneous compression and learning, which operates in the eigen space of the randomly initialized filters and learns to compactly represent the network as it trains from scratch. This eliminates the usual two-step process of having to first train the network, and then compressing it afterwards. To learn the sparse representations we enforce group L1 regularization on the linear combination weights of eigen filters. This results in the recombined filters which have low rank and can be readily compressed with standard pruning and low rank approximation methods. Moreover we show that the L1 norm of the linear combination weights can be used as a proxy for the filter importance for pruning. We demonstrate the effectiveness of our method by applying it to several CNN architectures, and show that our method directly achieves the best compression with competitive performance accuracy as compared to state of the art methods for compressing pre-trained networks.

It is well known that deep convolutional neural networks (CNNs) generalize well over large number of classes when ample training data is available. However, training with smaller datasets does not always achieve robust performance. In such cases, we show that using analytically derived filters in the lowest layer enables a network to achieve better performance than learning from scratch using a relatively small dataset. These class-agnostic filters represent the underlying manifold of the data space, and also generalize to new or unknown classes which may occur on the same manifold. This directly enables new classes to be learned with very few images by simply fine-tuning the final few layers of the network. We illustrate the advantages of our method using the publicly available set of infra-red images of vehicular ground targets. We compare a simple CNN trained using our method with transfer learning performed using the VGG-16 network, and show that when the number of training images is limited, the proposed approach not only achieves better results on the trained classes, but also outperforms a standard network for learning a new object class.

The detection of small moving targets in cluttered infrared imagery remains a difficult and challenging task. Conventional image subtraction techniques with frame-to-frame registration yield very high false alarm rates. Furthermore, state of the art deep convolutional neural networks (DCNNs) such as YOLO and Mask R-CNN also do not work well for this application. We show however, that it is possible to train a CNN to detect moving targets in a stack of stabilized images by maximizing a target to clutter ratio (TCR) metric. This metric has been previously used for detecting relatively large stationary targets in single images, but not for the purposes of finding small moving targets using multiple frames. Referred to as moving target indicator network (MTINet), the proposed network does not rely on image subtraction, but instead uses depth-wise convolution to learn inter-frame temporal dependencies. We compare the performance of the MTINet to state of the art DCNNs and a statistical anomaly detection algorithm, and propose a combined approach that offers the benefits of both data-driven learning and statistical analysis.

The detection of vehicular targets in infra-red imagery is a challenging task, both due to the relatively few pixels on target and the false alarms produced by the surrounding terrain clutter. It has been previously shown [1] that a relatively simple network (known as TCRNet) can outperform conventional deep CNNs for such applications by maximizing a target to clutter ratio (TCR) metric. In this paper, we introduce a new form of the network (referred to as TCRNet-2) that further improves the performance by first processing target and clutter information in two parallel channels and then combining them to optimize the TCR metric. We also show that the overall performance can be considerably improved by boosting the performance of a primary TCRNet-2 detector, with a secondary network that enhances discrimination between targets and clutter in the false alarm space of the primary network. We analyze the performance of the proposed networks using a publicly available data set of infra-red images of targets in natural terrain. It is shown that the TCRNet-2 and its boosted version yield considerably better performance than the original TCRNet over a wide range of distances, in both day and night conditions.

We propose an efficient and straightforward method for compressing deep convolutional neural networks (CNNs) that uses basis filters to represent the convolutional layers, and optimizes the performance of the compressed network directly in the basis space. Specifically, any spatial convolution layer of the CNN can be replaced by two successive convolution layers: the first is a set of three-dimensional orthonormal basis filters, followed by a layer of one-dimensional filters that represents the original spatial filters in the basis space. We jointly fine-tune both the basis and the filter representation to directly mitigate any performance loss due to the truncation. Generality of the proposed approach is demonstrated by applying it to several well known deep CNN architectures and data sets for image classification and object detection. We also present the execution time and power usage at different compression levels on the Xavier Jetson AGX processor.

The detection of targets in infra-red imagery is a challenging problem which involves locating small targets in heavily cluttered environments while maintaining a low false alarm rate. We propose a network that optimizes a “target to clutter ratio” (TCR) metric defined as the ratio of the output energies produced by the network in response to targets and clutter. We show that for target detection, it is advantageous to analytically derive the first layer of a CNN to maximize the TCR metric, and then train the rest of the network to optimize the same cost function. We evaluate the performance of the resulting network using a public domain MWIR data set released by the US Army's Night Vision Laboratories, and compare it to the state-of-the-art detectors such as Faster RCNN and Yolo-v3. Referred to as the TCRNet, the proposed network demonstrates state of the art results with greater than 30% improvement in probability of detection while reducing the false alarm rate by more than a factor of 2 when compared to these leading methods. Ablation studies also show that the proposed approach and metric are superior to learning the entire network from scratch, or using conventional regression metrics such as the mean square error (MSE).

The low resolution of infrared (IR) images makes feature extraction for classification of a challenging work. Learning-based methods, therefore, are preferred to be used on such raw imagery. In this article, in order to avoid difficulties in feature extraction, a novel multitask extension of the widely used sparse-representation-classification (SRC) method is proposed in both single and multiview set-ups. That is, the test sample could be a single IR image or images from different views. In both single-view and multiview scenarios, we try to employ collaborative spike and slab priors. This is because the traditional sparsity-inducing measures such as the l0-row pseudonorm makes it hard to capture the sparse structure of the coefficient matrix when expanded in terms of a training dictionary, and the priors are proved to be able to capture fairly general sparse structures. Furthermore, a joint prior and sparse coefficient estimation method (JPCEM) is proposed for the first time in this article in order to alleviate the need to handpick prior parameters required before classification. Multiple experiments are conducted on a synthetic Comanche Forward Looking IR (FLIR) Automatic Target Recognition (ATR) database collected by Army Research Lab and a challenging mid-wave IR (MWIR) image ATR database made available by the U.S. Army Night Vision and Electronic Sensors Directorate. The final results substantiate the merits of the proposed JPCEM through comparisons with other state-of-the-art methods, including both the ones based on SRC and the ones constructed using deep learning frameworks.

Support vector machine (SVM) classifiers are popular in many computer vision tasks. In most of them, the SVM classifier assumes that the object to be classified is centered in the query image, which might not always be valid, e.g., when locating and classifying a particular class of vehicles in a large scene. In this paper, we introduce a new classifier called Maximum Margin Correlation Filter (MMCF), which, while exhibiting the good generalization capabilities of SVM classifiers, is also capable of localizing objects of interest, thereby avoiding the need for image centering as is usually required in SVM classifiers. In other words, MMCF can simultaneously localize and classify objects of interest. We test the efficacy of the proposed classifier on three different tasks: vehicle recognition, eye localization, and face classification. We demonstrate that MMCF outperforms SVM classifiers as well as well known correlation filters.

We propose a contention based reservation MAC protocol for a collaborative sensing scenario involving a set of surveillance UAVs communicating with a hub. Data transmission rights are secured via distributed contention access between the UAV nodes for K available mini-slots. Each node chooses a slot with probability p independently of the others and is only allowed one attempt in a frame. We investigate the optimal choice of p which maximizes the one-shot probability of success(or alternately, the expected number of successes), beta, over the K minislots. We show that beta is prone to local maxima and develop an empirical formula based on goodness-of-fit which matches the numerically obtained values closely.

Correlation methods are becoming increasingly attractive tools for image recognition and location. This renewed interest in correlation methods is spurred by the availability of high-speed image processors and the emergence of correlation filter designs that can optimize relevant figures of merit. In this paper, a new correlation filter design method is presented that allows one to optimally tradeoff among potentially conflicting correlation output performance criteria while achieving desired correlation peak value behavior in response to in-plane rotation of input images. Such controlled in-plane rotation response is useful in image analysis and pattern recognition applications where the sensor follows a pre-arranged trajectory while imaging an object. Since this new correlation filter design is based on circular harmonic function (CHF) theory, we refer to the resulting filters as optimal tradeoff circular harmonic function (OTCHF) filters. Underlying theory, OTCHF filter design method, and illustrative numerical results are presented.

Terrain-delimitation is an important component of wide-area-surveillance with applications to battlefield terrain and agricultural terrain. Recently a statistical method was proposed by Mahalanobis and Singh [1] to design spatial filters to recognize and discriminate between various textures. We extend the technique for delimitation-through-texture-discrimination in SAR images. Spatial correlation filters are used for texture distinction. The filters are implementable as optical (or digital) correlators for fast real-time texture recognition without segmentation. The filter coefficients are determined via eigenvector analysis. Examples will be given to illustrate the proposed scheme for terrain-discrimination in SAR images.>