Shashank Agnihotri

Jan Philip Walter, Shashank Agnihotri, Margret Keuper

AI-generated image detection is a moving-target problem: detectors trained on one generator often fail when a new generator appears, and only a few labeled examples are available. We study a simple image-to-table formulation for this regime, where each image is encoded by a frozen DINOv3 backbone, its CLS feature is reduced to a 500-dimensional structured row with PCA, and TabPFN performs real/fake classification by in-context tabular inference rather than task-specific classifier training. This turns fake-image detection into low-data structured prediction over learned visual features, making detector adaptation depend on the labeled context set instead of gradient-based fine-tuning. On GenImage, LATTE, a recent state-of-the-art detector, remains stronger when many labeled samples from all generators are available, by 7.4% in the largest pooled setting, but DINOv3-PCA-TabPFN is stronger in the practically important low-data regime, outperforming LATTE by up to 8.2%, and in transfer settings where the detector must generalize from one generator to another. These results position tabular foundation models as a strong complementary adaptation mechanism for image forensics, shifting adaptation from detector retraining to lightweight in-context updates with a small labeled set of examples. Code URL: https://github.com/jpwalter30/Towards-Generalizable-Detection-of-AI-Generated-Images

Shashank Agnihotri, Steffen Jung, Margret Keuper

While neural networks allow highly accurate predictions in many tasks, their lack of robustness towards even slight input perturbations often hampers their deployment. Adversarial attacks such as the seminal projected gradient descent (PGD) offer an effective means to evaluate a model's robustness and dedicated solutions have been proposed for attacks on semantic segmentation or optical flow estimation. While they attempt to increase the attack's efficiency, a further objective is to balance its effect, so that it acts on the entire image domain instead of isolated point-wise predictions. This often comes at the cost of optimization stability and thus efficiency. Here, we propose CosPGD, an attack that encourages more balanced errors over the entire image domain while increasing the attack's overall efficiency. To this end, CosPGD leverages a simple alignment score computed from any pixel-wise prediction and its target to scale the loss in a smooth and fully differentiable way. It leads to efficient evaluations of a model's robustness for semantic segmentation as well as regression models (such as optical flow, disparity estimation, or image restoration), and it allows it to outperform the previous SotA attack on semantic segmentation. We provide code for the CosPGD algorithm and example usage at https://github.com/shashankskagnihotri/cospgd.

Hendrik Sommerhoff, Shashank Agnihotri, Mohamed Saleh et al.

The success of deep learning is frequently described as the ability to train all parameters of a network on a specific application in an end-to-end fashion. Yet, several design choices on the camera level, including the pixel layout of the sensor, are considered as pre-defined and fixed, and high resolution, regular pixel layouts are considered to be the most generic ones in computer vision and graphics, treating all regions of an image as equally important. While several works have considered non-uniform, \eg, hexagonal or foveated, pixel layouts in hardware and image processing, the layout has not been integrated into the end-to-end learning paradigm so far. In this work, we present the first truly end-to-end trained imaging pipeline that optimizes the size and distribution of pixels on the imaging sensor jointly with the parameters of a given neural network on a specific task. We derive an analytic, differentiable approach for the sensor layout parameterization that allows for task-specific, local varying pixel resolutions. We present two pixel layout parameterization functions: rectangular and curvilinear grid shapes that retain a regular topology. We provide a drop-in module that approximates sensor simulation given existing high-resolution images to directly connect our method with existing deep learning models. We show that network predictions benefit from learnable pixel layouts for two different downstream tasks, classification and semantic segmentation.

Shashank Agnihotri, Julia Grabinski, Margret Keuper

Pixel-wise predictions are required in a wide variety of tasks such as image restoration, image segmentation, or disparity estimation. Common models involve several stages of data resampling, in which the resolution of feature maps is first reduced to aggregate information and then increased to generate a high-resolution output. Previous works have shown that resampling operations are subject to artifacts such as aliasing. During downsampling, aliases have been shown to compromise the prediction stability of image classifiers. During upsampling, they have been leveraged to detect generated content. Yet, the effect of aliases during upsampling has not yet been discussed w.r.t. the stability and robustness of pixel-wise predictions. While falling under the same term (aliasing), the challenges for correct upsampling in neural networks differ significantly from those during downsampling: when downsampling, some high frequencies can not be correctly represented and have to be removed to avoid aliases. However, when upsampling for pixel-wise predictions, we actually require the model to restore such high frequencies that can not be encoded in lower resolutions. The application of findings from signal processing is therefore a necessary but not a sufficient condition to achieve the desirable output. In contrast, we find that the availability of large spatial context during upsampling allows to provide stable, high-quality pixel-wise predictions, even when fully learning all filter weights.

Shashank Agnihotri, Kanchana Vaishnavi Gandikota, Julia Grabinski et al.

Following their success in visual recognition tasks, Vision Transformers(ViTs) are being increasingly employed for image restoration. As a few recent works claim that ViTs for image classification also have better robustness properties, we investigate whether the improved adversarial robustness of ViTs extends to image restoration. We consider the recently proposed Restormer model, as well as NAFNet and the "Baseline network" which are both simplified versions of a Restormer. We use Projected Gradient Descent (PGD) and CosPGD, a recently proposed adversarial attack tailored to pixel-wise prediction tasks for our robustness evaluation. Our experiments are performed on real-world images from the GoPro dataset for image deblurring. Our analysis indicates that contrary to as advocated by ViTs in image classification works, these models are highly susceptible to adversarial attacks. We attempt to improve their robustness through adversarial training. While this yields a significant increase in robustness for Restormer, results on other networks are less promising. Interestingly, the design choices in NAFNet and Baselines, which were based on iid performance, and not on robust generalization, seem to be at odds with the model robustness. Thus, we investigate this further and find a fix.

Mishal Fatima, Shashank Agnihotri, Kanchana Vaishnavi Gandikota et al.

Most vision models are trained on RGB images processed through ISP pipelines optimized for human perception, which can discard sensor-level information useful for machine reasoning. RAW images preserve unprocessed scene data, enabling models to leverage richer cues for both object detection and object description, capturing fine-grained details, spatial relationships, and contextual information often lost in processed images. To support research in this domain, we introduce RAWDet-7, a large-scale dataset of ~25k training and 7.6k test RAW images collected across diverse cameras, lighting conditions, and environments, densely annotated for seven object categories following MS-COCO and LVIS conventions. In addition, we provide object-level descriptions derived from the corresponding high-resolution sRGB images, facilitating the study of object-level information preservation under RAW image processing and low-bit quantization. The dataset allows evaluation under simulated 4-bit, 6-bit, and 8-bit quantization, reflecting realistic sensor constraints, and provides a benchmark for studying detection performance, description quality & detail, and generalization in low-bit RAW image processing. Dataset & code upon acceptance.

Abhipsa Basu, Mohana Singh, Shashank Agnihotri et al.

Text-to-image (T2I) models are rapidly gaining popularity, yet their outputs often lack geographical diversity, reinforce stereotypes, and misrepresent regions. Given their broad reach, it is critical to rigorously evaluate how these models portray the world. Existing diversity metrics either rely on curated datasets or focus on surface-level visual similarity, limiting interpretability. We introduce GeoDiv, a framework leveraging large language and vision-language models to assess geographical diversity along two complementary axes: the Socio-Economic Visual Index (SEVI), capturing economic and condition-related cues, and the Visual Diversity Index (VDI), measuring variation in primary entities and backgrounds. Applied to images generated by models such as Stable Diffusion and FLUX.1-dev across $10$ entities and $16$ countries, GeoDiv reveals a consistent lack of diversity and identifies fine-grained attributes where models default to biased portrayals. Strikingly, depictions of countries like India, Nigeria, and Colombia are disproportionately impoverished and worn, reflecting underlying socio-economic biases. These results highlight the need for greater geographical nuance in generative models. GeoDiv provides the first systematic, interpretable framework for measuring such biases, marking a step toward fairer and more inclusive generative systems. Project page: https://abhipsabasu.github.io/geodiv

Shashank Agnihotri, David Schader, Nico Sharei et al.

Deep learning (DL) models are widely used in real-world applications but remain vulnerable to distribution shifts, especially due to weather and lighting changes. Collecting diverse real-world data for testing the robustness of DL models is resource-intensive, making synthetic corruptions an attractive alternative for robustness testing. However, are synthetic corruptions a reliable proxy for real-world corruptions? To answer this, we conduct the largest benchmarking study on semantic segmentation models, comparing performance on real-world corruptions and synthetic corruptions datasets. Our results reveal a strong correlation in mean performance, supporting the use of synthetic corruptions for robustness evaluation. We further analyze corruption-specific correlations, providing key insights to understand when synthetic corruptions succeed in representing real-world corruptions. Open-source Code: https://github.com/shashankskagnihotri/benchmarking_robustness/tree/segmentation_david/semantic_segmentation

Paul Gavrikov, Shashank Agnihotri, Margret Keuper et al.

Not all learnable parameters (e.g., weights) contribute equally to a neural network's decision function. In fact, entire layers' parameters can sometimes be reset to random values with little to no impact on the model's decisions. We revisit earlier studies that examined how architecture and task complexity influence this phenomenon and ask: is this phenomenon also affected by how we train the model? We conducted experimental evaluations on a diverse set of ImageNet-1k classification models to explore this, keeping the architecture and training data constant but varying the training pipeline. Our findings reveal that the training method strongly influences which layers become critical to the decision function for a given task. For example, improved training regimes and self-supervised training increase the importance of early layers while significantly under-utilizing deeper layers. In contrast, methods such as adversarial training display an opposite trend. Our preliminary results extend previous findings, offering a more nuanced understanding of the inner mechanics of neural networks. Code: https://github.com/paulgavrikov/layer_criticality

Long Phan, Alice Gatti, Ziwen Han et al. · amazon-science, apple-ml

Benchmarks are important tools for tracking the rapid advancements in large language model (LLM) capabilities. However, benchmarks are not keeping pace in difficulty: LLMs now achieve over 90\% accuracy on popular benchmarks like MMLU, limiting informed measurement of state-of-the-art LLM capabilities. In response, we introduce Humanity's Last Exam (HLE), a multi-modal benchmark at the frontier of human knowledge, designed to be the final closed-ended academic benchmark of its kind with broad subject coverage. HLE consists of 2,500 questions across dozens of subjects, including mathematics, humanities, and the natural sciences. HLE is developed globally by subject-matter experts and consists of multiple-choice and short-answer questions suitable for automated grading. Each question has a known solution that is unambiguous and easily verifiable, but cannot be quickly answered via internet retrieval. State-of-the-art LLMs demonstrate low accuracy and calibration on HLE, highlighting a significant gap between current LLM capabilities and the expert human frontier on closed-ended academic questions. To inform research and policymaking upon a clear understanding of model capabilities, we publicly release HLE at https://lastexam.ai.

Shashank Agnihotri, Jonas Jakubassa, Priyam Dey et al.

Open-weight LLMs can be modified at inference time with simple activation edits, which raises a practical question for safety: do common safety interventions like refusal training or metatag training survive such edits? We study model abliteration, a lightweight projection technique designed to remove refusal-sensitive directions, and conduct a controlled evaluation across a granular sequence of Safety Pretraining checkpoints for SmolLM2-1.7B, alongside widely used open baselines. For each of 20 systems, original and abliterated, we issue 100 prompts with balanced harmful and harmless cases, classify responses as **Refusal** or **Non-Refusal** using multiple judges, and validate judge fidelity on a small human-labeled subset. We also probe whether models can identify refusal in their own outputs. Our study produces a checkpoint-level characterization of which data-centric safety components remain robust under abliteration, quantifies how judge selection influences evaluation outcomes, and outlines a practical protocol for integrating inference-time edits into safety assessments. Code: https://github.com/shashankskagnihotri/safety_pretraining.

Mishal Fatima, Shashank Agnihotri, Marius Bock et al.

Most pattern recognition models are developed on pre-proce\-ssed data. In computer vision, for instance, RGB images processed through image signal processing (ISP) pipelines designed to cater to human perception are the most frequent input to image analysis networks. However, many modern vision tasks operate without a human in the loop, raising the question of whether such pre-processing is optimal for automated analysis. Similarly, human activity recognition (HAR) on body-worn sensor data commonly takes normalized floating-point data arising from a high-bit analog-to-digital converter (ADC) as an input, despite such an approach being highly inefficient in terms of data transmission, significantly affecting the battery life of wearable devices. In this work, we target low-bandwidth and energy-constrained settings where sensors are limited to low-bit-depth capture. We propose $γ$-Quant, i.e.~the task-specific learning of a non-linear quantization for pattern recognition. We exemplify our approach on raw-image object detection as well as HAR of wearable data, and demonstrate that raw data with a learnable quantization using as few as 4-bits can perform on par with the use of raw 12-bit data. All code to reproduce our experiments is publicly available via https://github.com/Mishalfatima/Gamma-Quant

Marcel Kleinmann, Shashank Agnihotri, Margret Keuper

Faithfulness and interpretability are essential for deploying deep neural networks (DNNs) in safety-critical domains such as medical imaging. B-cos networks offer a promising solution by replacing standard linear layers with a weight-input alignment mechanism, producing inherently interpretable, class-specific explanations without post-hoc methods. While maintaining diagnostic performance competitive with state-of-the-art DNNs, standard B-cos models suffer from severe aliasing artifacts in their explanation maps, making them unsuitable for clinical use where clarity is essential. In this work, we address these limitations by introducing anti-aliasing strategies using FLCPooling (FLC) and BlurPool (BP) to significantly improve explanation quality. Our experiments on chest X-ray datasets demonstrate that the modified $\text{B-cos}_\text{FLC}$ and $\text{B-cos}_\text{BP}$ preserve strong predictive performance while providing faithful and artifact-free explanations suitable for clinical application in multi-class and multi-label settings. Code available at: GitHub repository (url: https://github.com/mkleinma/B-cos-medical-paper).

Nicolas Poggi, Shashank Agnihotri, Margret Keuper

Terahertz (THz) imaging enables non-invasive analysis for applications such as security screening and material classification, but effective image classification remains challenging due to limited annotations, low resolution, and visual ambiguity. We introduce In-Context Learning (ICL) with Vision-Language Models (VLMs) as a flexible, interpretable alternative that requires no fine-tuning. Using a modality-aligned prompting framework, we adapt two open-weight VLMs to the THz domain and evaluate them under zero-shot and one-shot settings. Our results show that ICL improves classification and interpretability in low-data regimes. This is the first application of ICL-enhanced VLMs to THz imaging, offering a promising direction for resource-constrained scientific domains. Code: \href{https://github.com/Nicolas-Poggi/Project_THz_Classification/tree/main}{GitHub repository}.

Shashank Agnihotri, David Schader, Jonas Jakubassa et al.

Reliability and generalization in deep learning are predominantly studied in the context of image classification. Yet, real-world applications in safety-critical domains involve a broader set of semantic tasks, such as semantic segmentation and object detection, which come with a diverse set of dedicated model architectures. To facilitate research towards robust model design in segmentation and detection, our primary objective is to provide benchmarking tools regarding robustness to distribution shifts and adversarial manipulations. We propose the benchmarking tools SEMSEGBENCH and DETECBENCH, along with the most extensive evaluation to date on the reliability and generalization of semantic segmentation and object detection models. In particular, we benchmark 76 segmentation models across four datasets and 61 object detectors across two datasets, evaluating their performance under diverse adversarial attacks and common corruptions. Our findings reveal systematic weaknesses in state-of-the-art models and uncover key trends based on architecture, backbone, and model capacity. SEMSEGBENCH and DETECBENCH are open-sourced in our GitHub repository (https://github.com/shashankskagnihotri/benchmarking_reliability_generalization) along with our complete set of total 6139 evaluations. We anticipate the collected data to foster and encourage future research towards improved model reliability beyond classification.

Shashank Agnihotri, Amaan Ansari, Annika Dackermann et al.

Deep learning (DL) has surpassed human performance on standard benchmarks, driving its widespread adoption in computer vision tasks. One such task is disparity estimation, estimating the disparity between matching pixels in stereo image pairs, which is crucial for safety-critical applications like medical surgeries and autonomous navigation. However, DL-based disparity estimation methods are highly susceptible to distribution shifts and adversarial attacks, raising concerns about their reliability and generalization. Despite these concerns, a standardized benchmark for evaluating the robustness of disparity estimation methods remains absent, hindering progress in the field. To address this gap, we introduce DispBench, a comprehensive benchmarking tool for systematically assessing the reliability of disparity estimation methods. DispBench evaluates robustness against synthetic image corruptions such as adversarial attacks and out-of-distribution shifts caused by 2D Common Corruptions across multiple datasets and diverse corruption scenarios. We conduct the most extensive performance and robustness analysis of disparity estimation methods to date, uncovering key correlations between accuracy, reliability, and generalization. Open-source code for DispBench: https://github.com/shashankskagnihotri/benchmarking_robustness/tree/disparity_estimation/final/disparity_estimation

Jenny Schmalfuss, Victor Oei, Lukas Mehl et al.

Standard benchmarks for optical flow, scene flow, and stereo vision algorithms generally focus on model accuracy rather than robustness to image corruptions like noise or rain. Hence, the resilience of models to such real-world perturbations is largely unquantified. To address this, we present RobustSpring, a comprehensive dataset and benchmark for evaluating robustness to image corruptions for optical flow, scene flow, and stereo models. RobustSpring applies 20 different image corruptions, including noise, blur, color changes, quality degradations, and weather distortions, in a time-, stereo-, and depth-consistent manner to the high-resolution Spring dataset, creating a suite of 20,000 corrupted images that reflect challenging conditions. RobustSpring enables comparisons of model robustness via a new corruption robustness metric. Integration with the Spring benchmark enables public two-axis evaluations of both accuracy and robustness. We benchmark a curated selection of initial models, observing that accurate models are not necessarily robust and that robustness varies widely by corruption type. RobustSpring is a new computer vision benchmark that treats robustness as a first-class citizen to foster models that combine accuracy with resilience. It will be available at https://spring-benchmark.org.

Eyad Alshami, Shashank Agnihotri, Bernt Schiele et al.

It has been observed that deep neural networks (DNNs) often use both genuine as well as spurious features. In this work, we propose "Amending Inherent Interpretability via Self-Supervised Masking" (AIM), a simple yet interestingly effective method that promotes the network's utilization of genuine features over spurious alternatives without requiring additional annotations. In particular, AIM uses features at multiple encoding stages to guide a self-supervised, sample-specific feature-masking process. As a result, AIM enables the training of well-performing and inherently interpretable models that faithfully summarize the decision process. We validate AIM across a diverse range of challenging datasets that test both out-of-distribution generalization and fine-grained visual understanding. These include general-purpose classification benchmarks such as ImageNet100, HardImageNet, and ImageWoof, as well as fine-grained classification datasets such as Waterbirds, TravelingBirds, and CUB-200. AIM demonstrates significant dual benefits: interpretability improvements, as measured by the Energy Pointing Game (EPG) score, and accuracy gains over strong baselines. These consistent gains across domains and architectures provide compelling evidence that AIM promotes the use of genuine and meaningful features that directly contribute to improved generalization and human-aligned interpretability.

Lorena Stracke, Lia Nimmermann, Shashank Agnihotri et al.

Inspired by the human visual system's mechanisms for contrast enhancement and color-opponency, we explore biologically motivated input preprocessing for robust semantic segmentation. By applying Difference-of-Gaussians (DoG) filtering to RGB, grayscale, and opponent-color channels, we enhance local contrast without modifying model architecture or training. Evaluations on Cityscapes, ACDC, and Dark Zurich show that such preprocessing maintains in-distribution performance while improving robustness to adverse conditions like night, fog, and snow. As this processing is model-agnostic and lightweight, it holds potential for integration into imaging pipelines, enabling imaging systems to deliver task-ready, robust inputs for downstream vision models in safety-critical environments.

Shashank Agnihotri, Julia Grabinski, Janis Keuper et al.

Image restoration networks are usually comprised of an encoder and a decoder, responsible for aggregating image content from noisy, distorted data and to restore clean, undistorted images, respectively. Data aggregation as well as high-resolution image generation both usually come at the risk of involving aliases, i.e.~standard architectures put their ability to reconstruct the model input in jeopardy to reach high PSNR values on validation data. The price to be paid is low model robustness. In this work, we show that simply providing alias-free paths in state-of-the-art reconstruction transformers supports improved model robustness at low costs on the restoration performance. We do so by proposing BOA-Restormer, a transformer-based image restoration model that executes downsampling and upsampling operations partly in the frequency domain to ensure alias-free paths along the entire model while potentially preserving all relevant high-frequency information.