Abhijit Mahalanobis

Surendar Devasundaram, Saber Latibari Banafsheh, Abhijit Mahalanobis

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

Daniel Brignac, Niels Lobo, Abhijit Mahalanobis

Continual learning seeks to enable deep learners to train on a series of tasks of unknown length without suffering from the catastrophic forgetting of previous tasks. One effective solution is replay, which involves storing few previous experiences in memory and replaying them when learning the current task. However, there is still room for improvement when it comes to selecting the most informative samples for storage and determining the optimal number of samples to be stored. This study aims to address these issues with a novel comparison of the commonly used reservoir sampling to various alternative population strategies and providing a novel detailed analysis of how to find the optimal number of stored samples.

Daniel Brignac, Fengwei Tian, Banafsheh Latibari et al.

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.

Alex Berian, Daniel Brignac, JhihYang Wu et al.

Geospatial imaging leverages data from diverse sensing modalities-such as EO, SAR, and LiDAR, ranging from ground-level drones to satellite views. These heterogeneous inputs offer significant opportunities for scene understanding but present challenges in interpreting geometry accurately, particularly in the absence of precise ground truth data. To address this, we propose CrossModalityDiffusion, a modular framework designed to generate images across different modalities and viewpoints without prior knowledge of scene geometry. CrossModalityDiffusion employs modality-specific encoders that take multiple input images and produce geometry-aware feature volumes that encode scene structure relative to their input camera positions. The space where the feature volumes are placed acts as a common ground for unifying input modalities. These feature volumes are overlapped and rendered into feature images from novel perspectives using volumetric rendering techniques. The rendered feature images are used as conditioning inputs for a modality-specific diffusion model, enabling the synthesis of novel images for the desired output modality. In this paper, we show that jointly training different modules ensures consistent geometric understanding across all modalities within the framework. We validate CrossModalityDiffusion's capabilities on the synthetic ShapeNet cars dataset, demonstrating its effectiveness in generating accurate and consistent novel views across multiple imaging modalities and perspectives.

Alex Beriand, JhihYang Wu, Daniel Brignac et al.

Novel view synthesis (NVS) has advanced with generative modeling, enabling photorealistic image generation. In few-shot NVS, where only a few input views are available, existing methods often assume equal importance for all input views relative to the target, leading to suboptimal results. We address this limitation by introducing a camera-weighting mechanism that adjusts the importance of source views based on their relevance to the target. We propose two approaches: a deterministic weighting scheme leveraging geometric properties like Euclidean distance and angular differences, and a cross-attention-based learning scheme that optimizes view weighting. Additionally, models can be further trained with our camera-weighting scheme to refine their understanding of view relevance and enhance synthesis quality. This mechanism is adaptable and can be integrated into various NVS algorithms, improving their ability to synthesize high-quality novel views. Our results demonstrate that adaptive view weighting enhances accuracy and realism, offering a promising direction for improving NVS.

Najmeh Nazari, Banafsheh Saber Latibari, Elahe Hosseini et al.

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.

Banafsheh Saber Latibari, Najmeh Nazari, Hossein Sayadi et al.

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, Brain 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 the 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.

Daniel Brignac, Abhijit Mahalanobis

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.

Natnael Daba, Bruce McIntosh, Abhijit Mahalanobis

Conventional classifiers are trained and evaluated using balanced data sets in which all classes are equally present. Classifiers are now trained on large data sets such as ImageNet, and are now able to classify hundreds (if not thousands) of different classes. On one hand, it is desirable to train such general-purpose classifier on a very large number of classes so that it performs well regardless of the settings in which it is deployed. On the other hand, it is unlikely that all classes known to the classifier will occur in every deployment scenario, or that they will occur with the same prior probability. In reality, only a relatively small subset of the known classes may be present in a particular setting or environment. For example, a classifier will encounter mostly animals if its deployed in a zoo or for monitoring wildlife, aircraft and service vehicles at an airport, or various types of automobiles and commercial vehicles if it is used for monitoring traffic. Furthermore, the exact class priors are generally unknown and can vary over time. In this paper, we explore different methods for estimating the class priors based on the output of the classifier itself. We then show that incorporating the estimated class priors in the overall decision scheme enables the classifier to increase its run-time accuracy in the context of its deployment scenario.

Maliha Arif, Calvin Yong, Abhijit Mahalanobis

Even though convolutional neural networks 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 negatively affect their performance. We propose a new two-step training procedure called split training to reduce this bias in CNNs on both Infrared imagery and RGB data. Our split training procedure has two steps: using MSE loss first train the layers of the network on images with background to match the activations of the same network when it is trained using images without background; then with these layers frozen, train the rest of the network with cross-entropy loss to classify the objects. Our training method outperforms the traditional training procedure in both a simple CNN architecture, and deep CNNs like VGG and Densenet which use lots of hardware resources, and learns to mimic human vision which focuses more on shape and structure than background with higher accuracy.

Muhammad Tayyab, Fahad Ahmad Khan, Abhijit Mahalanobis

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.

Maliha Arif, Abhijit Mahalanobis

We propose a novel study of generating unseen arbitrary viewpoints for infrared imagery in the non-linear feature subspace . Current methods use synthetic images and often result in blurry and distorted outputs. Our approach on the contrary understands the semantic information in natural images and encapsulates it such that our predicted unseen views possess good 3D representations. We further explore the non-linear feature subspace and conclude that our network does not operate in the Euclidean subspace but rather in the Riemannian subspace. It does not learn the geometric transformation for predicting the position of the pixel in the new image but rather learns the manifold. To this end, we use t-SNE visualisations to conduct a detailed analysis of our network and perform classification of generated images as a low-shot learning task.

Shashanka Venkataramanan, Kuan-Chuan Peng, Rajat Vikram Singh et al.

Anomaly localization is an important problem in computer vision which involves localizing anomalous regions within images with applications in industrial inspection, surveillance, and medical imaging. This task is challenging due to the small sample size and pixel coverage of the anomaly in real-world scenarios. Most prior works need to use anomalous training images to compute a class-specific threshold to localize anomalies. Without the need of anomalous training images, we propose Convolutional Adversarial Variational autoencoder with Guided Attention (CAVGA), which localizes the anomaly with a convolutional latent variable to preserve the spatial information. In the unsupervised setting, we propose an attention expansion loss where we encourage CAVGA to focus on all normal regions in the image. Furthermore, in the weakly-supervised setting we propose a complementary guided attention loss, where we encourage the attention map to focus on all normal regions while minimizing the attention map corresponding to anomalous regions in the image. CAVGA outperforms the state-of-the-art (SOTA) anomaly localization methods on MVTec Anomaly Detection (MVTAD), modified ShanghaiTech Campus (mSTC) and Large-scale Attention based Glaucoma (LAG) datasets in the unsupervised setting and when using only 2% anomalous images in the weakly-supervised setting. CAVGA also outperforms SOTA anomaly detection methods on the MNIST, CIFAR-10, Fashion-MNIST, MVTAD, mSTC and LAG datasets.

Muhammad Tayyab, Abhijit Mahalanobis

It is well known that Convolutional Neural Networks (CNNs) have significant redundancy in their filter weights. Various methods have been proposed in the literature to compress trained CNNs. These include techniques like pruning weights, filter quantization and representing filters in terms of a basis functions. Our approach falls in this latter class of strategies, but is distinct in that that we show both compressed learning and representation can be achieved without significant modifications of popular CNN architectures. Specifically, any convolution layer of the CNN is easily replaced by two successive convolution layers: the first is a set of fixed filters (that represent the knowledge space of the entire layer and do not change), which is followed by a layer of one-dimensional filters (that represent the learned knowledge in this space). For the pre-trained networks, the fixed layer is just the truncated eigen-decompositions of the original filters. The 1D filters are initialized as the weights of linear combination, but are fine-tuned to recover any performance loss due to the truncation. For training networks from scratch, we use a set of random orthogonal fixed filters (that never change), and learn the 1D weight vector directly from the labeled data. Our method substantially reduces i) the number of learnable parameters during training, and ii) the number of multiplication operations and filter storage requirements during implementation. It does so without requiring any special operators in the convolution layer, and extends to all known popular CNN architectures. We apply our method to four well known network architectures trained with three different data sets. Results show a consistent reduction in i) the number of operations by up to a factor of 5, and ii) number of learnable parameters by up to a factor of 18, with less than 3% drop in performance on the CIFAR100 dataset.