Baidu USA LLC

Haichao Zhang

Jianyu Wang

Wei Xu

Haonan Yu

Described herein are embodiments for fast training of adversarially robust models against adversarial attacks. Embodiments for model training by perturbing both the image and the label, which may be referred to as Bilateral Adversarial Training (BAT), are disclosed. To generate the adversarial label, one or more closed-form heuristic solutions are derived. One-step targeted attack is used with the target label being the most confusing class. It is shown in various experiments that random start and the most confusing target attack effectively prevent the label leaking and gradient masking problem. Coupled with the adversarial label part, embodiments of presented models significantly improve the state-of-the-art results. Experiments on one or more computationally challenging dataset also demonstrate the effectiveness of the presented BAT method embodiments.

Systems and methods for fast training of more robust models against adversarial attacks

Described herein are embodiments for a feature-scattering-based adversarial training approach for improving model robustness against adversarial attacks. Conventional adversarial training approaches leverage a supervised scheme, either targeted or non-targeted in generating attacks for training, which typically suffer from issues such as label leaking as noted in recent works. Embodiments of the disclosed approach generate adversarial images for training through feature scattering in the latent space, which is unsupervised in nature and avoids label leaking. More importantly, the presented approaches generate perturbed images in a collaborative fashion, taking the inter-sample relationships into consideration. Extensive experiments on different datasets compared with state-of-the-art approaches demonstrate the effectiveness of the presented embodiments.

Systems and methods for defense against adversarial attacks using feature scattering-based adversarial training

Described herein are systems and methods for grounded natural language learning in an interactive setting. In embodiments, during a learning process, an agent learns natural language by interacting with a teacher and learning from feedback, thus learning and improving language skills while taking part in the conversation. In embodiments, a model is used to incorporate both imitation and reinforcement by leveraging jointly sentence and reward feedback from the teacher. Various experiments are conducted to validate the effectiveness of a model embodiment.

Listen, interact, and talk: learning to speak via interaction

Described herein are systems and methods for interactive language acquisition with one-shot concept learning through a conversational environment. Supervised language learning is limited by the ability of capturing mainly the statistics of training data, and is hardly adaptive to new scenarios or flexible for acquiring new knowledge without inefficient retraining or catastrophic forgetting. In one or more embodiments, conversational interaction serves as a natural interface both for language learning and for novel knowledge acquisition. Embodiments of a joint imitation and reinforcement approach are disclosed for grounded language learning through interactive conversation. An agent trained with this approach is able to actively acquire information by asking questions about novel objects and use the just-learned knowledge in subsequent conversations in a one-shot fashion. Results compared with other methods verified the effectiveness of embodiments disclosed herein.

Systems and methods for interactive language acquisition with one-shot visual concept learning through a conversational game

Described herein are embodiments for an approach to improve the robustness of an object detector against adversarial attacks. Existing attacks for object detectors and the impacts of individual task component on model robustness are systematically analyzed from a multi-task view of object detection. In one or more embodiments, a multi-task learning perspective of object detection is introduced and an asymmetric role of task losses is identified. One or more embodiments of an adversarial training method for robust object detection are presented to leverage the multiple sources of attacks for improving the robustness of object detection models.

Systems and methods for adversarially robust object detection

Described herein are embodiments for joint adversarial training methods that incorporate both spatial transformation-based and pixel-value based attacks for improving image model robustness. Embodiments of a spatial transformation-based attack with an explicit notion of budgets are disclosed and embodiments of a practical methodology for efficient spatial attack generation are also disclosed. Furthermore, both pixel and spatial attacks are integrated into embodiments of a generation model and the complementary strengths of each other are leveraged for improving the overall model robustness. Extensive experimental results on several benchmark datasets compared with state-of-the-art methods verified the effectiveness of the presented method.

Systems and methods for joint adversarial training by incorporating both spatial and pixel attacks

Described herein are systems and methods for human-like language acquisition in a compositional framework to implement object recognition or navigation tasks. Embodiments include a method for a model to learn the input language in a grounded and compositional manner, such that after training the model is able to correctly execute zero-shot commands, which have either combination of words in the command never appeared before, and/or new object concepts learned from another task but never learned from navigation settings. In embodiments, a framework is trained end-to-end to learn simultaneously the visual representations of the environment, the syntax and semantics of the language, and outputs actions via an action module. In embodiments, the zero-shot learning capability of a framework results from its compositionality and modularity with parameter tying.

Growing community of inventors

Sunnyvale, CA, United States of America

Haichao Zhang