Aleksandr Petiushko

Jiawei Zhang, Xuan Yang, Taiqi Wang et al.

Traditional autonomous driving systems often struggle to connect high-level reasoning with low-level control, leading to suboptimal and sometimes unsafe behaviors. Recent advances in multimodal large language models (MLLMs), which process both visual and textual data, offer an opportunity to unify perception and reasoning. However, effectively embedding precise safety knowledge into MLLMs for autonomous driving remains a significant challenge. To address this, we propose SafeAuto, a framework that enhances MLLM-based autonomous driving by incorporating both unstructured and structured knowledge. First, we introduce a Position-Dependent Cross-Entropy (PDCE) loss to improve low-level control signal predictions when values are represented as text. Second, to explicitly integrate safety knowledge, we develop a reasoning component that translates traffic rules into first-order logic (e.g., "red light $\implies$ stop") and embeds them into a probabilistic graphical model (e.g., Markov Logic Network) to verify predicted actions using recognized environmental attributes. Additionally, our Multimodal Retrieval-Augmented Generation (RAG) model leverages video, control signals, and environmental attributes to learn from past driving experiences. Integrating PDCE, MLN, and Multimodal RAG, SafeAuto outperforms existing baselines across multiple datasets, enabling more accurate, reliable, and safer autonomous driving. The code is available at https://github.com/AI-secure/SafeAuto.

Zihan Zhou, Jonathan Booher, Khashayar Rohanimanesh et al.

Safe reinforcement learning tasks are a challenging domain despite being very common in the real world. The widely adopted CMDP model constrains the risks in expectation, which makes room for dangerous behaviors in long-tail states. In safety-critical domains, such behaviors could lead to disastrous outcomes. To address this issue, we first describe the problem with a stronger Uniformly Constrained MDP (UCMDP) model where we impose constraints on all reachable states; we then propose Objective Suppression, a novel method that adaptively suppresses the task reward maximizing objectives according to a safety critic, as a solution to the Lagrangian dual of a UCMDP. We benchmark Objective Suppression in two multi-constraint safety domains, including an autonomous driving domain where any incorrect behavior can lead to disastrous consequences. On the driving domain, we evaluate on open source and proprietary data and evaluate transfer to a real autonomous fleet. Empirically, we demonstrate that our proposed method, when combined with existing safe RL algorithms, can match the task reward achieved by baselines with significantly fewer constraint violations.

Andrey Kaznacheev, Matvey Mikhalchuk, Andrey Kuznetsov et al.

Adversarial attacks on face recognition (FR) systems pose a significant security threat, yet most are confined to the digital domain or require white-box access. We introduce GaP (Gaussian Patch), a novel method to generate a universal, physically transferable adversarial patch under a strict black-box setting. Our approach uses a query-efficient, zero-order greedy algorithm to iteratively construct a symmetric, grayscale pattern for the forehead. The patch is optimized by successively adding Gaussian blobs, guided only by the cosine similarity scores from a surrogate FR model to maximally degrade identity recognition. We demonstrate that with approximately 10,000 queries to a black-box ArcFace model, the resulting GaP achieves a high attack success rate in both digital and real-world physical tests. Critically, the attack shows strong transferability, successfully deceiving an entirely unseen FaceNet model. Our work highlights a practical and severe vulnerability, proving that robust, transferable attacks can be crafted with limited knowledge of the target system.

Anton Razzhigaev, Matvey Mikhalchuk, Klim Kireev et al.

Reconstructing facial images from black-box recognition models poses a significant privacy threat. While many methods require access to embeddings, we address the more challenging scenario of model inversion using only similarity scores. This paper introduces DarkerBB, a novel approach that reconstructs color faces by performing zero-order optimization within a PCA-derived eigenface space. Despite this highly limited information, experiments on LFW, AgeDB-30, and CFP-FP benchmarks demonstrate that DarkerBB achieves state-of-the-art verification accuracies in the similarity-only setting, with competitive query efficiency.

Mikhail Pautov, Olesya Kuznetsova, Nurislam Tursynbek et al.

Randomized smoothing is considered to be the state-of-the-art provable defense against adversarial perturbations. However, it heavily exploits the fact that classifiers map input objects to class probabilities and do not focus on the ones that learn a metric space in which classification is performed by computing distances to embeddings of classes prototypes. In this work, we extend randomized smoothing to few-shot learning models that map inputs to normalized embeddings. We provide analysis of Lipschitz continuity of such models and derive robustness certificate against $\ell_2$-bounded perturbations that may be useful in few-shot learning scenarios. Our theoretical results are confirmed by experiments on different datasets.

Daria Bakshandaeva, Denis Dimitrov, Vladimir Arkhipkin et al.

Supporting the current trend in the AI community, we present the AI Journey 2021 Challenge called Fusion Brain, the first competition which is targeted to make the universal architecture which could process different modalities (in this case, images, texts, and code) and solve multiple tasks for vision and language. The Fusion Brain Challenge combines the following specific tasks: Code2code Translation, Handwritten Text recognition, Zero-shot Object Detection, and Visual Question Answering. We have created datasets for each task to test the participants' submissions on it. Moreover, we have collected and made publicly available a new handwritten dataset in both English and Russian, which consists of 94,128 pairs of images and texts. We also propose a multimodal and multitask architecture - a baseline solution, in the center of which is a frozen foundation model and which has been trained in Fusion mode along with Single-task mode. The proposed Fusion approach proves to be competitive and more energy-efficient compared to the task-specific one.

Mikhail Pautov, Nurislam Tursynbek, Marina Munkhoeva et al.

In safety-critical machine learning applications, it is crucial to defend models against adversarial attacks -- small modifications of the input that change the predictions. Besides rigorously studied $\ell_p$-bounded additive perturbations, recently proposed semantic perturbations (e.g. rotation, translation) raise a serious concern on deploying ML systems in real-world. Therefore, it is important to provide provable guarantees for deep learning models against semantically meaningful input transformations. In this paper, we propose a new universal probabilistic certification approach based on Chernoff-Cramer bounds that can be used in general attack settings. We estimate the probability of a model to fail if the attack is sampled from a certain distribution. Our theoretical findings are supported by experimental results on different datasets.

Alexander Ivanov, Gleb Nosovskiy, Alexey Chekunov et al.

Manifold hypothesis states that data points in high-dimensional space actually lie in close vicinity of a manifold of much lower dimension. In many cases this hypothesis was empirically verified and used to enhance unsupervised and semi-supervised learning. Here we present new approach to manifold hypothesis checking and underlying manifold dimension estimation. In order to do it we use two very different methods simultaneously - one geometric, another probabilistic - and check whether they give the same result. Our geometrical method is a modification for sparse data of a well-known box-counting algorithm for Minkowski dimension calculation. The probabilistic method is new. Although it exploits standard nearest neighborhood distance, it is different from methods which were previously used in such situations. This method is robust, fast and includes special preliminary data transformation. Experiments on real datasets show that the suggested approach based on two methods combination is powerful and effective.

Nikita Muravev, Aleksandr Petiushko

Currently the most popular method of providing robustness certificates is randomized smoothing where an input is smoothed via some probability distribution. We propose a novel approach to randomized smoothing over multiplicative parameters. Using this method we construct certifiably robust classifiers with respect to a gamma correction perturbation and compare the result with classifiers obtained via other smoothing distributions (Gaussian, Laplace, uniform). The experiments show that asymmetrical Rayleigh distribution allows to obtain better certificates for some values of perturbation parameters. To the best of our knowledge it is the first work concerning certified robustness against the multiplicative gamma correction transformation and the first to study effects of asymmetrical distributions in randomized smoothing.

Anton Razzhigaev, Klim Kireev, Igor Udovichenko et al.

Several methods for inversion of face recognition models were recently presented, attempting to reconstruct a face from deep templates. Although some of these approaches work in a black-box setup using only face embeddings, usually, on the end-user side, only similarity scores are provided. Therefore, these algorithms are inapplicable in such scenarios. We propose a novel approach that allows reconstructing the face querying only similarity scores of the black-box model. While our algorithm operates in a more general setup, experiments show that it is query efficient and outperforms the existing methods.

Maksim Dzabraev, Maksim Kalashnikov, Stepan Komkov et al.

We present a new state-of-the-art on the text to video retrieval task on MSRVTT and LSMDC benchmarks where our model outperforms all previous solutions by a large margin. Moreover, state-of-the-art results are achieved with a single model on two datasets without finetuning. This multidomain generalisation is achieved by a proper combination of different video caption datasets. We show that training on different datasets can improve test results of each other. Additionally we check intersection between many popular datasets and found that MSRVTT has a significant overlap between the test and the train parts, and the same situation is observed for ActivityNet.

Fedor Pavutnitskiy, Sergei O. Ivanov, Evgeny Abramov et al.

The knowledge that data lies close to a particular submanifold of the ambient Euclidean space may be useful in a number of ways. For instance, one may want to automatically mark any point far away from the submanifold as an outlier or to use the geometry to come up with a better distance metric. Manifold learning problems are often posed in a very high dimension, e.g. for spaces of images or spaces of words. Today, with deep representation learning on the rise in areas such as computer vision and natural language processing, many problems of this kind may be transformed into problems of moderately high dimension, typically of the order of hundreds. Motivated by this, we propose a manifold learning technique suitable for moderately high dimension and large datasets. The manifold is learned from the training data in the form of an intersection of quadric hypersurfaces -- simple but expressive objects. At test time, this manifold can be used to introduce a computationally efficient outlier score for arbitrary new data points and to improve a given similarity metric by incorporating the learned geometric structure into it.

Nurislam Tursynbek, Aleksandr Petiushko, Ivan Oseledets

Differential privacy (DP) is a gold-standard concept of measuring and guaranteeing privacy in data analysis. It is well-known that the cost of adding DP to deep learning model is its accuracy. However, it remains unclear how it affects robustness of the model. Standard neural networks are not robust to different input perturbations: either adversarial attacks or common corruptions. In this paper, we empirically observe an interesting trade-off between privacy and robustness of neural networks. We experimentally demonstrate that networks, trained with DP, in some settings might be even more vulnerable in comparison to non-private versions. To explore this, we extensively study different robustness measurements, including FGSM and PGD adversaries, distance to linear decision boundaries, curvature profile, and performance on a corrupted dataset. Finally, we study how the main ingredients of differentially private neural networks training, such as gradient clipping and noise addition, affect (decrease and increase) the robustness of the model.

Stepan Komkov, Maksim Dzabraev, Aleksandr Petiushko

The construction of models for video action classification progresses rapidly. However, the performance of those models can still be easily improved by ensembling with the same models trained on different modalities (e.g. Optical flow). Unfortunately, it is computationally expensive to use several modalities during inference. Recent works examine the ways to integrate advantages of multi-modality into a single RGB-model. Yet, there is still a room for improvement. In this paper, we explore the various methods to embed the ensemble power into a single model. We show that proper initialization, as well as mutual modality learning, enhances single-modality models. As a result, we achieve state-of-the-art results in the Something-Something-v2 benchmark.

Nurislam Tursynbek, Aleksandr Petiushko, Ivan Oseledets

The brittleness of deep image classifiers to small adversarial input perturbations has been extensively studied in the last several years. However, the main objective of existing perturbations is primarily limited to change the correctly predicted Top-1 class by an incorrect one, which does not intend to change the Top-k prediction. In many digital real-world scenarios Top-k prediction is more relevant. In this work, we propose a fast and accurate method of computing Top-k adversarial examples as a simple multi-objective optimization. We demonstrate its efficacy and performance by comparing it to other adversarial example crafting techniques. Moreover, based on this method, we propose Top-k Universal Adversarial Perturbations, image-agnostic tiny perturbations that cause the true class to be absent among the Top-k prediction for the majority of natural images. We experimentally show that our approach outperforms baseline methods and even improves existing techniques of finding Universal Adversarial Perturbations.

Stepan Komkov, Aleksandr Petiushko

In this paper we propose a novel easily reproducible technique to attack the best public Face ID system ArcFace in different shooting conditions. To create an attack, we print the rectangular paper sticker on a common color printer and put it on the hat. The adversarial sticker is prepared with a novel algorithm for off-plane transformations of the image which imitates sticker location on the hat. Such an approach confuses the state-of-the-art public Face ID model LResNet100E-IR, ArcFace@ms1m-refine-v2 and is transferable to other Face ID models.