Adversarial attacks on traffic sign classification models were among the first successfully tried in the real world. Since then, the research in this area has been mainly restricted to repeating baseline models, such as LISA-CNN or GTSRB-CNN, and similar experiment settings, including white and black patches on traffic signs. In this work, we decouple model architectures from the datasets and evaluate on further generic models to make a fair comparison. Furthermore, we compare two attack settings, inconspicuous and visible, which are usually regarded without direct comparison. Our results show that standard baselines like LISA-CNN or GTSRB-CNN are significantly more susceptible than the generic ones. We, therefore, suggest evaluating new attacks on a broader spectrum of baselines in the future. Our code is available at \url{https://github.com/KASTEL-MobilityLab/attacks-on-traffic-sign-recognition/}.
Despite ongoing efforts to defend neural classifiers from adversarial attacks, they remain vulnerable, especially to unseen attacks. In contrast, humans are difficult to be cheated by subtle manipulations, since we make judgments only based on essential factors. Inspired by this observation, we attempt to model label generation with essential label-causative factors and incorporate label-non-causative factors to assist data generation. For an adversarial example, we aim to discriminate the perturbations as non-causative factors and make predictions only based on the label-causative factors. Concretely, we propose a casual diffusion model (CausalDiff) that adapts diffusion models for conditional data generation and disentangles the two types of casual factors by learning towards a novel casual information bottleneck objective. Empirically, CausalDiff has significantly outperformed state-of-the-art defense methods on various unseen attacks, achieving an average robustness of 86.39% (+4.01%) on CIFAR-10, 56.25% (+3.13%) on CIFAR-100, and 82.62% (+4.93%) on GTSRB (German Traffic Sign Recognition Benchmark). The code is available at https://github.com/CAS-AISafetyBasicResearchGroup/CausalDiff.
Traffic sign is a critical map feature for navigation and traffic control. Nevertheless, current methods for traffic sign recognition rely on traditional deep learning models, which typically suffer from significant performance degradation considering the variations in data distribution across different regions. In this paper, we propose TSCLIP, a robust fine-tuning approach with the contrastive language-image pre-training (CLIP) model for worldwide cross-regional traffic sign recognition. We first curate a cross-regional traffic sign benchmark dataset by combining data from ten different sources. Then, we propose a prompt engineering scheme tailored to the characteristics of traffic signs, which involves specific scene descriptions and corresponding rules to generate targeted text descriptions for optimizing the model training process. During the TSCLIP fine-tuning process, we implement adaptive dynamic weight ensembling (ADWE) to seamlessly incorporate outcomes from each training iteration with the zero-shot CLIP model. This approach ensures that the model retains its ability to generalize while acquiring new knowledge about traffic signs. Our method surpasses conventional classification benchmark models in cross-regional traffic sign evaluations, and it achieves state-of-the-art performance compared to existing CLIP fine-tuning techniques. To the best knowledge of authors, TSCLIP is the first contrastive language-image model used for the worldwide cross-regional traffic sign recognition task. The project website is available at: https://github.com/guoyangzhao/TSCLIP.
Traffic signs play a critical role in road safety and traffic management for autonomous driving systems. Accurate traffic sign classification is essential but challenging due to real-world complexities like adversarial examples and occlusions. To address these issues, binary neural networks offer promise in constructing classifiers suitable for resource-constrained devices. In our previous work, we proposed high-accuracy BNN models for traffic sign recognition, focusing on compact size for limited computation and energy resources. To evaluate their local robustness, this paper introduces a set of benchmark problems featuring layers that challenge state-of-the-art verification tools. These layers include binarized convolutions, max pooling, batch normalization, fully connected. The difficulty of the verification problem is given by the high number of network parameters (905k - 1.7 M), of the input dimension (2.7k-12k), and of the number of regions (43) as well by the fact that the neural networks are not sparse. The proposed BNN models and local robustness properties can be checked at https://github.com/ChristopherBrix/vnncomp2023_benchmarks/tree/main/benchmarks/traffic_signs_recognition. The results of the 4th International Verification of Neural Networks Competition (VNN-COMP'23) revealed the fact that 4, out of 7, solvers can handle many of our benchmarks randomly selected (minimum is 6, maximum is 36, out of 45). Surprisingly, tools output also wrong results or missing counterexample (ranging from 1 to 4). Currently, our focus lies in exploring the possibility of achieving a greater count of solved instances by extending the allotted time (previously set at 8 minutes). Furthermore, we are intrigued by the reasons behind the erroneous outcomes provided by the tools for certain benchmarks.
All current backdoor attacks on deep learning (DL) models fall under the category of a vertical class backdoor (VCB) -- class-dependent. In VCB attacks, any sample from a class activates the implanted backdoor when the secret trigger is present. Existing defense strategies overwhelmingly focus on countering VCB attacks, especially those that are source-class-agnostic. This narrow focus neglects the potential threat of other simpler yet general backdoor types, leading to false security implications. This study introduces a new, simple, and general type of backdoor attack coined as the horizontal class backdoor (HCB) that trivially breaches the class dependence characteristic of the VCB, bringing a fresh perspective to the community. HCB is now activated when the trigger is presented together with an innocuous feature, regardless of class. For example, the facial recognition model misclassifies a person who wears sunglasses with a smiling innocuous feature into the targeted person, such as an administrator, regardless of which person. The key is that these innocuous features are horizontally shared among classes but are only exhibited by partial samples per class. Extensive experiments on attacking performance across various tasks, including MNIST, facial recognition, traffic sign recognition, object detection, and medical diagnosis, confirm the high efficiency and effectiveness of the HCB. We rigorously evaluated the evasiveness of the HCB against a series of eleven representative countermeasures, including Fine-Pruning (RAID 18'), STRIP (ACSAC 19'), Neural Cleanse (Oakland 19'), ABS (CCS 19'), Februus (ACSAC 20'), NAD (ICLR 21'), MNTD (Oakland 21'), SCAn (USENIX SEC 21'), MOTH (Oakland 22'), Beatrix (NDSS 23'), and MM-BD (Oakland 24'). None of these countermeasures prove robustness, even when employing a simplistic trigger, such as a small and static white-square patch.
We study the problem of certifying the robustness of Bayesian neural networks (BNNs) to adversarial input perturbations. Given a compact set of input points $T \subseteq \mathbb{R}^m$ and a set of output points $S \subseteq \mathbb{R}^n$, we define two notions of robustness for BNNs in an adversarial setting: probabilistic robustness and decision robustness. Probabilistic robustness is the probability that for all points in $T$ the output of a BNN sampled from the posterior is in $S$. On the other hand, decision robustness considers the optimal decision of a BNN and checks if for all points in $T$ the optimal decision of the BNN for a given loss function lies within the output set $S$. Although exact computation of these robustness properties is challenging due to the probabilistic and non-convex nature of BNNs, we present a unified computational framework for efficiently and formally bounding them. Our approach is based on weight interval sampling, integration, and bound propagation techniques, and can be applied to BNNs with a large number of parameters, and independently of the (approximate) inference method employed to train the BNN. We evaluate the effectiveness of our methods on various regression and classification tasks, including an industrial regression benchmark, MNIST, traffic sign recognition, and airborne collision avoidance, and demonstrate that our approach enables certification of robustness and uncertainty of BNN predictions.
Existing datasets for autonomous driving (AD) often lack diversity and long-range capabilities, focusing instead on 360{\deg} perception and temporal reasoning. To address this gap, we introduce Zenseact Open Dataset (ZOD), a large-scale and diverse multimodal dataset collected over two years in various European countries, covering an area 9x that of existing datasets. ZOD boasts the highest range and resolution sensors among comparable datasets, coupled with detailed keyframe annotations for 2D and 3D objects (up to 245m), road instance/semantic segmentation, traffic sign recognition, and road classification. We believe that this unique combination will facilitate breakthroughs in long-range perception and multi-task learning. The dataset is composed of Frames, Sequences, and Drives, designed to encompass both data diversity and support for spatio-temporal learning, sensor fusion, localization, and mapping. Frames consist of 100k curated camera images with two seconds of other supporting sensor data, while the 1473 Sequences and 29 Drives include the entire sensor suite for 20 seconds and a few minutes, respectively. ZOD is the only large-scale AD dataset released under a permissive license, allowing for both research and commercial use. More information, and an extensive devkit, can be found at https://zod.zenseact.com
Although there are many datasets for traffic sign classification, there are few datasets collected for traffic sign recognition and few of them obtain enough instances especially for training a model with the deep learning method. The deep learning method is almost the only way to train a model for real-world usage that covers various highly similar classes compared with the traditional way such as through color, shape, etc. Also, for some certain sign classes, their sign meanings were destined to can't get enough instances in the dataset. To solve this problem, we purpose a unique data augmentation method for the traffic sign recognition dataset that takes advantage of the standard of the traffic sign. We called it TSR dataset augmentation. We based on the benchmark Tsinghua-Tencent 100K (TT100K) dataset to verify the unique data augmentation method. we performed the method on four main iteration version datasets based on the TT100K dataset and the experimental results showed our method is efficacious. The iteration version datasets based on TT100K, data augmentation method source code and the training results introduced in this paper are publicly available.
Traffic signs support road safety and managing the flow of traffic, hence are an integral part of any vision system for autonomous driving. While the use of deep learning is well-known in traffic signs classification due to the high accuracy results obtained using convolutional neural networks (CNNs) (state of the art is 99.46\%), little is known about binarized neural networks (BNNs). Compared to CNNs, BNNs reduce the model size and simplify convolution operations and have shown promising results in computationally limited and energy-constrained devices which appear in the context of autonomous driving. This work presents a bottom-up approach for architecturing BNNs by studying characteristics of the constituent layers. These constituent layers (binarized convolutional layers, max pooling, batch normalization, fully connected layers) are studied in various combinations and with different values of kernel size, number of filters and of neurons by using the German Traffic Sign Recognition Benchmark (GTSRB) for training. As a result, we propose BNNs architectures which achieve more than $90\%$ for GTSRB (the maximum is $96.45\%$) and an average greater than $80\%$ (the maximum is $88.99\%$) considering also the Belgian and Chinese datasets for testing. The number of parameters of these architectures varies from 100k to less than 2M. The accompanying material of this paper is publicly available at https://github.com/apostovan21/BinarizedNeuralNetwork.
Vision transformers have been demonstrated to yield state-of-the-art results on a variety of computer vision tasks using attention-based networks. However, research works in transformers mostly do not investigate robustness/accuracy trade-off, and they still struggle to handle adversarial perturbations. In this paper, we explore the robustness of vision transformers against adversarial perturbations and try to enhance their robustness/accuracy trade-off in white box attack settings. To this end, we propose Locality iN Locality (LNL) transformer model. We prove that the locality introduction to LNL contributes to the robustness performance since it aggregates local information such as lines, edges, shapes, and even objects. In addition, to further improve the robustness performance, we encourage LNL to extract training signal from the moments (a.k.a., mean and standard deviation) and the normalized features. We validate the effectiveness and generality of LNL by achieving state-of-the-art results in terms of accuracy and robustness metrics on German Traffic Sign Recognition Benchmark (GTSRB) and Canadian Institute for Advanced Research (CIFAR-10). More specifically, for traffic sign classification, the proposed LNL yields gains of 1.1% and ~35% in terms of clean and robustness accuracy compared to the state-of-the-art studies.
Adversarial patch is an important form of real-world adversarial attack that brings serious risks to the robustness of deep neural networks. Previous methods generate adversarial patches by either optimizing their perturbation values while fixing the pasting position or manipulating the position while fixing the patch's content. This reveals that the positions and perturbations are both important to the adversarial attack. For that, in this paper, we propose a novel method to simultaneously optimize the position and perturbation for an adversarial patch, and thus obtain a high attack success rate in the black-box setting. Technically, we regard the patch's position, the pre-designed hyper-parameters to determine the patch's perturbations as the variables, and utilize the reinforcement learning framework to simultaneously solve for the optimal solution based on the rewards obtained from the target model with a small number of queries. Extensive experiments are conducted on the Face Recognition (FR) task, and results on four representative FR models show that our method can significantly improve the attack success rate and query efficiency. Besides, experiments on the commercial FR service and physical environments confirm its practical application value. We also extend our method to the traffic sign recognition task to verify its generalization ability.
Real-time machine learning object detection algorithms are often found within autonomous vehicle technology and depend on quality datasets. It is essential that these algorithms work correctly in everyday conditions as well as under strong sun glare. Reports indicate glare is one of the two most prominent environment-related reasons for crashes. However, existing datasets, such as the Laboratory for Intelligent & Safe Automobiles Traffic Sign (LISA) Dataset and the German Traffic Sign Recognition Benchmark, do not reflect the existence of sun glare at all. This paper presents the GLARE (GLARE is available at: https://github.com/NicholasCG/GLARE_Dataset ) traffic sign dataset: a collection of images with U.S-based traffic signs under heavy visual interference by sunlight. GLARE contains 2,157 images of traffic signs with sun glare, pulled from 33 videos of dashcam footage of roads in the United States. It provides an essential enrichment to the widely used LISA Traffic Sign dataset. Our experimental study shows that although several state-of-the-art baseline architectures have demonstrated good performance on traffic sign detection in conditions without sun glare in the past, they performed poorly when tested against GLARE (e.g., average mAP0.5:0.95 of 19.4). We also notice that current architectures have better detection when trained on images of traffic signs in sun glare performance (e.g., average mAP0.5:0.95 of 39.6), and perform best when trained on a mixture of conditions (e.g., average mAP0.5:0.95 of 42.3).
Machine Learning~(ML) has provided promising results in recent years across different applications and domains. However, in many cases, qualities such as reliability or even safety need to be ensured. To this end, one important aspect is to determine whether or not ML components are deployed in situations that are appropriate for their application scope. For components whose environments are open and variable, for instance those found in autonomous vehicles, it is therefore important to monitor their operational situation to determine its distance from the ML components' trained scope. If that distance is deemed too great, the application may choose to consider the ML component outcome unreliable and switch to alternatives, e.g. using human operator input instead. SafeML is a model-agnostic approach for performing such monitoring, using distance measures based on statistical testing of the training and operational datasets. Limitations in setting SafeML up properly include the lack of a systematic approach for determining, for a given application, how many operational samples are needed to yield reliable distance information as well as to determine an appropriate distance threshold. In this work, we address these limitations by providing a practical approach and demonstrate its use in a well known traffic sign recognition problem, and on an example using the CARLA open-source automotive simulator.
Estimating the risk level of adversarial examples is essential for safely deploying machine learning models in the real world. One popular approach for physical-world attacks is to adopt the "sticker-pasting" strategy, which however suffers from some limitations, including difficulties in access to the target or printing by valid colors. A new type of non-invasive attacks emerged recently, which attempt to cast perturbation onto the target by optics based tools, such as laser beam and projector. However, the added optical patterns are artificial but not natural. Thus, they are still conspicuous and attention-grabbed, and can be easily noticed by humans. In this paper, we study a new type of optical adversarial examples, in which the perturbations are generated by a very common natural phenomenon, shadow, to achieve naturalistic and stealthy physical-world adversarial attack under the black-box setting. We extensively evaluate the effectiveness of this new attack on both simulated and real-world environments. Experimental results on traffic sign recognition demonstrate that our algorithm can generate adversarial examples effectively, reaching 98.23% and 90.47% success rates on LISA and GTSRB test sets respectively, while continuously misleading a moving camera over 95% of the time in real-world scenarios. We also offer discussions about the limitations and the defense mechanism of this attack.
Discriminative self-supervised learning allows training models on any random group of internet images, and possibly recover salient information that helps differentiate between the images. Applied to ImageNet, this leads to object centric features that perform on par with supervised features on most object-centric downstream tasks. In this work, we question if using this ability, we can learn any salient and more representative information present in diverse unbounded set of images from across the globe. To do so, we train models on billions of random images without any data pre-processing or prior assumptions about what we want the model to learn. We scale our model size to dense 10 billion parameters to avoid underfitting on a large data size. We extensively study and validate our model performance on over 50 benchmarks including fairness, robustness to distribution shift, geographical diversity, fine grained recognition, image copy detection and many image classification datasets. The resulting model, not only captures well semantic information, it also captures information about artistic style and learns salient information such as geolocations and multilingual word embeddings based on visual content only. More importantly, we discover that such model is more robust, more fair, less harmful and less biased than supervised models or models trained on object centric datasets such as ImageNet.
We describe an iterative active-learning algorithm to recognise rare traffic signs. A standard ResNet is trained on a training set containing only a single sample of the rare class. We demonstrate that by sorting the samples of a large, unlabeled set by the estimated probability of belonging to the rare class, we can efficiently identify samples from the rare class. This works despite the fact that this estimated probability is usually quite low. A reliable active-learning loop is obtained by labeling these candidate samples, including them in the training set, and iterating the procedure. Further, we show that we get similar results starting from a single synthetic sample. Our results are important as they indicate a straightforward way of improving traffic-sign recognition for automated driving systems. In addition, they show that we can make use of the information hidden in low confidence outputs, which is usually ignored.
To assess the vulnerability of deep learning in the physical world, recent works introduce adversarial patches and apply them on different tasks. In this paper, we propose another kind of adversarial patch: the Meaningful Adversarial Sticker, a physically feasible and stealthy attack method by using real stickers existing in our life. Unlike the previous adversarial patches by designing perturbations, our method manipulates the sticker's pasting position and rotation angle on the objects to perform physical attacks. Because the position and rotation angle are less affected by the printing loss and color distortion, adversarial stickers can keep good attacking performance in the physical world. Besides, to make adversarial stickers more practical in real scenes, we conduct attacks in the black-box setting with the limited information rather than the white-box setting with all the details of threat models. To effectively solve for the sticker's parameters, we design the Region based Heuristic Differential Evolution Algorithm, which utilizes the new-found regional aggregation of effective solutions and the adaptive adjustment strategy of the evaluation criteria. Our method is comprehensively verified in the face recognition and then extended to the image retrieval and traffic sign recognition. Extensive experiments show the proposed method is effective and efficient in complex physical conditions and has a good generalization for different tasks.
For visual object recognition tasks, the illumination variations can cause distinct changes in object appearance and thus confuse the deep neural network based recognition models. Especially for some rare illumination conditions, collecting sufficient training samples could be time-consuming and expensive. To solve this problem, in this paper we propose a novel neural network architecture called Separating-Illumination Network (Sill-Net). Sill-Net learns to separate illumination features from images, and then during training we augment training samples with these separated illumination features in the feature space. Experimental results demonstrate that our approach outperforms current state-of-the-art methods in several object classification benchmarks.
Real world traffic sign recognition is an important step towards building autonomous vehicles, most of which highly dependent on Deep Neural Networks (DNNs). Recent studies demonstrated that DNNs are surprisingly susceptible to adversarial examples. Many attack methods have been proposed to understand and generate adversarial examples, such as gradient based attack, score based attack, decision based attack, and transfer based attacks. However, most of these algorithms are ineffective in real-world road sign attack, because (1) iteratively learning perturbations for each frame is not realistic for a fast moving car and (2) most optimization algorithms traverse all pixels equally without considering their diverse contribution. To alleviate these problems, this paper proposes the targeted attention attack (TAA) method for real world road sign attack. Specifically, we have made the following contributions: (1) we leverage the soft attention map to highlight those important pixels and skip those zero-contributed areas - this also helps to generate natural perturbations, (2) we design an efficient universal attack that optimizes a single perturbation/noise based on a set of training images under the guidance of the pre-trained attention map, (3) we design a simple objective function that can be easily optimized, (4) we evaluate the effectiveness of TAA on real world data sets. Experimental results validate that the TAA method improves the attack successful rate (nearly 10%) and reduces the perturbation loss (about a quarter) compared with the popular RP2 method. Additionally, our TAA also provides good properties, e.g., transferability and generalization capability. We provide code and data to ensure the reproducibility: https://github.com/AdvAttack/RoadSignAttack.
Research into adversarial examples (AE) has developed rapidly, yet static adversarial patches are still the main technique for conducting attacks in the real world, despite being obvious, semi-permanent and unmodifiable once deployed. In this paper, we propose Short-Lived Adversarial Perturbations (SLAP), a novel technique that allows adversaries to realize physically robust real-world AE by using a light projector. Attackers can project a specifically crafted adversarial perturbation onto a real-world object, transforming it into an AE. This allows the adversary greater control over the attack compared to adversarial patches: (i) projections can be dynamically turned on and off or modified at will, (ii) projections do not suffer from the locality constraint imposed by patches, making them harder to detect. We study the feasibility of SLAP in the self-driving scenario, targeting both object detector and traffic sign recognition tasks, focusing on the detection of stop signs. We conduct experiments in a variety of ambient light conditions, including outdoors, showing how in non-bright settings the proposed method generates AE that are extremely robust, causing misclassifications on state-of-the-art networks with up to 99% success rate for a variety of angles and distances. We also demostrate that SLAP-generated AE do not present detectable behaviours seen in adversarial patches and therefore bypass SentiNet, a physical AE detection method. We evaluate other defences including an adaptive defender using adversarial learning which is able to thwart the attack effectiveness up to 80% even in favourable attacker conditions.