Current remote-sensing interpretation models often focus on a single task such as detection, segmentation, or caption. However, the task-specific designed models are unattainable to achieve the comprehensive multi-level interpretation of images. The field also lacks support for multi-task joint interpretation datasets. In this paper, we propose Panoptic Perception, a novel task and a new fine-grained dataset (FineGrip) to achieve a more thorough and universal interpretation for RSIs. The new task, 1) integrates pixel-level, instance-level, and image-level information for universal image perception, 2) captures image information from coarse to fine granularity, achieving deeper scene understanding and description, and 3) enables various independent tasks to complement and enhance each other through multi-task learning. By emphasizing multi-task interactions and the consistency of perception results, this task enables the simultaneous processing of fine-grained foreground instance segmentation, background semantic segmentation, and global fine-grained image captioning. Concretely, the FineGrip dataset includes 2,649 remote sensing images, 12,054 fine-grained instance segmentation masks belonging to 20 foreground things categories, 7,599 background semantic masks for 5 stuff classes and 13,245 captioning sentences. Furthermore, we propose a joint optimization-based panoptic perception model. Experimental results on FineGrip demonstrate the feasibility of the panoptic perception task and the beneficial effect of multi-task joint optimization on individual tasks. The dataset will be publicly available.
Recent research has proposed approaches that modify speech to defend against gender inference attacks. The goal of these protection algorithms is to control the availability of information about a speaker's gender, a privacy-sensitive attribute. Currently, the common practice for developing and testing gender protection algorithms is "neural-on-neural", i.e., perturbations are generated and tested with a neural network. In this paper, we propose to go beyond this practice to strengthen the study of gender protection. First, we demonstrate the importance of testing gender inference attacks that are based on speech features historically developed by speech scientists, alongside the conventionally used neural classifiers. Next, we argue that researchers should use speech features to gain insight into how protective modifications change the speech signal. Finally, we point out that gender-protection algorithms should be compared with novel "vocal adversaries", human-executed voice adaptations, in order to improve interpretability and enable before-the-mic protection.
Perturbative availability poisoning (PAP) adds small changes to images to prevent their use for model training. Current research adopts the belief that practical and effective approaches to countering such poisons do not exist. In this paper, we argue that it is time to abandon this belief. We present extensive experiments showing that 12 state-of-the-art PAP methods are vulnerable to Image Shortcut Squeezing (ISS), which is based on simple compression. For example, on average, ISS restores the CIFAR-10 model accuracy to $81.73\%$, surpassing the previous best preprocessing-based countermeasures by $37.97\%$ absolute. ISS also (slightly) outperforms adversarial training and has higher generalizability to unseen perturbation norms and also higher efficiency. Our investigation reveals that the property of PAP perturbations depends on the type of surrogate model used for poison generation, and it explains why a specific ISS compression yields the best performance for a specific type of PAP perturbation. We further test stronger, adaptive poisoning, and show it falls short of being an ideal defense against ISS. Overall, our results demonstrate the importance of considering various (simple) countermeasures to ensure the meaningfulness of analysis carried out during the development of availability poisons.
We introduce ShortcutGen, a new data poisoning attack that generates sample-dependent, error-minimizing perturbations by learning a generator. The key novelty of ShortcutGen is the use of a randomly-initialized discriminator, which provides spurious shortcuts needed for generating poisons. Different from recent, iterative methods, our ShortcutGen can generate perturbations with only one forward pass in a label-free manner, and compared to the only existing generative method, DeepConfuse, our ShortcutGen is faster and simpler to train while remaining competitive. We also demonstrate that integrating a simple augmentation strategy can further boost the robustness of ShortcutGen against early stopping, and combining augmentation and non-augmentation leads to new state-of-the-art results in terms of final validation accuracy, especially in the challenging, transfer scenario. Lastly, we speculate, through uncovering its working mechanism, that learning a more general representation space could allow ShortcutGen to work for unseen data.
Building a scalable and real-time recommendation system is vital for many businesses driven by time-sensitive customer feedback, such as short-videos ranking or online ads. Despite the ubiquitous adoption of production-scale deep learning frameworks like TensorFlow or PyTorch, these general-purpose frameworks fall short of business demands in recommendation scenarios for various reasons: on one hand, tweaking systems based on static parameters and dense computations for recommendation with dynamic and sparse features is detrimental to model quality; on the other hand, such frameworks are designed with batch-training stage and serving stage completely separated, preventing the model from interacting with customer feedback in real-time. These issues led us to reexamine traditional approaches and explore radically different design choices. In this paper, we present Monolith, a system tailored for online training. Our design has been driven by observations of our application workloads and production environment that reflects a marked departure from other recommendations systems. Our contributions are manifold: first, we crafted a collisionless embedding table with optimizations such as expirable embeddings and frequency filtering to reduce its memory footprint; second, we provide an production-ready online training architecture with high fault-tolerance; finally, we proved that system reliability could be traded-off for real-time learning. Monolith has successfully landed in the BytePlus Recommend product.
Strengthening the robustness of machine learning-based malware detectors against realistic evasion attacks remains one of the major obstacles for Android malware detection. To that end, existing work has focused on interpreting domain constraints of Android malware in the problem space, where problem-space realizable adversarial examples are generated. In this paper, we provide another promising way to achieve the same goal but based on interpreting the domain constraints in the feature space, where feature-space realizable adversarial examples are generated. Specifically, we present a novel approach to extracting feature-space domain constraints by learning meaningful feature dependencies from data, and applying them based on a novel robust feature space. Experimental results successfully demonstrate the effectiveness of our novel robust feature space in providing adversarial robustness for DREBIN, a state-of-the-art Android malware detector. For example, it can decrease the evasion rate of a realistic gradient-based attack by $96.4\%$ in a limited-knowledge (transfer) setting and by $13.8\%$ in a more challenging, perfect-knowledge setting. In addition, we show that directly using our learned domain constraints in the adversarial retraining framework leads to about $84\%$ improvement in a limited-knowledge setting, with up to $377\times$ faster implementation than using problem-space adversarial examples.
Recent work has shown that imperceptible perturbations can be applied to craft unlearnable examples (ULEs), i.e. images whose content cannot be used to improve a classifier during training. In this paper, we reveal the road that researchers should follow for understanding ULEs and improving ULEs as they were originally formulated (ULEOs). The paper makes four contributions. First, we show that ULEOs exploit color and, consequently, their effects can be mitigated by simple grayscale pre-filtering, without resorting to adversarial training. Second, we propose an extension to ULEOs, which is called ULEO-GrayAugs, that forces the generated ULEs away from channel-wise color perturbations by making use of grayscale knowledge and data augmentations during optimization. Third, we show that ULEOs generated using Multi-Layer Perceptrons (MLPs) are effective in the case of complex Convolutional Neural Network (CNN) classifiers, suggesting that CNNs suffer specific vulnerability to ULEs. Fourth, we demonstrate that when a classifier is trained on ULEOs, adversarial training will prevent a drop in accuracy measured both on clean images and on adversarial images. Taken together, our contributions represent a substantial advance in the state of art of unlearnable examples, but also reveal important characteristics of their behavior that must be better understood in order to achieve further improvements.
There is broad consensus among researchers studying adversarial examples that it is extremely difficult to achieve transferability of targeted attacks. Currently, existing research strives for transferability of targeted attacks by resorting to sophisticated losses and even massive training. In this paper, we take a second look at the transferability of targeted attacks and show that their difficulty has been overestimated due to a blind spot in the conventional evaluation procedures. Specifically, current work has unreasonably restricted attack optimization to a few iterations. Here, we show that targeted attacks converge slowly to optimal transferability and improve considerably when given more iterations. We also demonstrate that an attack that simply maximizes the target logit performs surprisingly well, remarkably surpassing more complex losses and even achieving performance comparable to the state of the art, which requires massive training with sophisticated loss. We provide further validation of our logit attack in a realistic ensemble setting and in a real-world attack against the Google Cloud Vision. The logit attack produces perturbations that reflect the target semantics, which we demonstrate allows us to create targeted universal adversarial perturbations without additional training images.
We propose Adversarial Color Enhancement (ACE), a novel approach to generating non-suspicious adversarial images by optimizing a color transformation within a parametric filter space. The filter we use approximates human-understandable color curve adjustment, constraining ACE with a single, continuous function. This property gives rise to a principled adversarial action space explicitly controlled by filter parameters. Existing color transformation attacks are not guided by a parametric space, and, consequently, additional pixel-related constraints such as regularization and sampling are necessary. These constraints make methodical analysis difficult. In this paper, we carry out a systematic robustness analysis of ACE from both the attack and defense perspectives by varying the bound of the color filter parameters. We investigate a general formulation of ACE and also a variant targeting particularly appealing color styles, as achieved with popular image filters. From the attack perspective, we provide extensive experiments on the vulnerability of image classifiers, but also explore the vulnerability of segmentation and aesthetic quality assessment algorithms, in both the white-box and black-box scenarios. From the defense perspective, more experiments provide insight into the stability of ACE against input transformation-based defenses and show the potential of adversarial training for improving model robustness against ACE.
E-commerce platforms provide their customers with ranked lists of recommended items matching the customers' preferences. Merchants on e-commerce platforms would like their items to appear as high as possible in the top-N of these ranked lists. In this paper, we demonstrate how unscrupulous merchants can create item images that artificially promote their products, improving their rankings. Recommender systems that use images to address the cold start problem are vulnerable to this security risk. We describe a new type of attack, Adversarial Item Promotion (AIP), that strikes directly at the core of Top-N recommenders: the ranking mechanism itself. Existing work on adversarial images in recommender systems investigates the implications of conventional attacks, which target deep learning classifiers. In contrast, our AIP attacks are embedding attacks that seek to push features representations in a way that fools the ranker (not a classifier) and directly lead to item promotion. We introduce three AIP attacks insider attack, expert attack, and semantic attack, which are defined with respect to three successively more realistic attack models. Our experiments evaluate the danger of these attacks when mounted against three representative visually-aware recommender algorithms in a framework that uses images to address cold start. We also evaluate two common defenses against adversarial images in the classification scenario and show that these simple defenses do not eliminate the danger of AIP attacks. In sum, we show that using images to address cold start opens recommender systems to potential threats with clear practical implications. To facilitate future research, we release an implementation of our attacks and defenses, which allows reproduction and extension.