Fudan University

Huawei Technologies Co., Limited

Yugang Jiang

Yanwei Fu

Xiangyang Xue

Yinda Zhang

Zhuwen Li

Shaoxiang Chen

Xingjun Ma

Nanyang Wang

Jie Liu

Dong Wang

Jingjing Chen

Yingbin Zheng

Shihao Zhao

Changmao Cheng

Linxi Jiang

Systems and methods for video backdoor attack include a trigger generation module for generating a universal adversarial trigger pattern specific to a task, an adversarial perturbation module for producing videos with manipulated features; and a poisoning and inference module for injecting the generated trigger into perturbed videos as poisoned samples for training; wherein the trigger pattern is patched and optimized on videos from all non-target classes but relabeled to a target class, and the trigger pattern is a universal adversarial trigger pattern generated by minimizing the cross-entropy loss.

System and method for video backdoor attack

A network for image processing is provided, and more particularly, for coarse-to-fine recognition of image processing. The network includes a shared convolution layer, and a first subnet and a second subnet both subsequent to the shared convolution layer; the first subnet comprises a first skipping module comprising one or more skip-dense blocks iteratively stacked with one or more transition layers, a first pooling layer subsequent to the first skipping module, and a first classification layer subsequent to the first pooling layer; the second subnet comprises a second skipping module comprising one or more skip-dense blocks iteratively stacked with one or more layers, a second pooling layer subsequent to the second skipping module, and a second classification layer subsequent to the second pooling layer; and wherein a skip-dense block of the second subnet is selected to guide a transition layer of the first subnet, and the level of the guiding skip-dense block is deeper than the level of the guided transition layer. This network is also related to a system and a method thereof.

Network, system and method for image processing

A system for activity localization in videos is described, comprising a visual concept detection module, which produces a plurality of first visual concept vectors each representing a probability of containing visual concepts for one of a plurality of sampled frames sampled from an input video; wherein each of the plurality of first visual concept vectors dot-product with a second visual concept vector extracted from a given query sentence, resulting a visual-semantic correlation score; a semantic activity proposal generation module, which generates semantic activity proposals by temporally grouping frames with a high visual-semantic correlation score; and a proposal evaluation and refinement module, which takes the semantic activity proposals, the visual concept vectors and the query sentence as input, and outputs alignment scores and refined boundaries for the proposals. The disclosure also relates to methods thereof.

System and method for localization of activities in videos

A method for generating black-box adversarial attacks on video recognition models is provided, comprising a) passing input video frames into a public image model, to obtain pixel-wise tentative perturbations; b) partitioning the tentative perturbations into tentative perturbation patches; c) estimating the rectification weight required for each patch, via querying the target video model; d) applying the patch-wise rectification weight on the patches, to obtain the rectified pixel-wise perturbations; e) applying one step projected gradient descent (PGD) perturbation on the input video, according to the rectified pixel-wise perturbations; and f) iteratively performing steps a)-e) until an attack succeeds or a query limit is reached. Systems and networks therefor are also provided.

Black-box adversarial attacks on videos

This invention is related to a network for generating 3D shape, including an image feature network, an initial ellipsoid mesh, and a cascaded mesh deformation network. The image feature network is a Visual Geometry Group Net (VGGN) containing five successive convolutional layer groups, and four pooling layers sandwiched by the five convolutional layer groups; and the cascaded mesh deformation network is a graph-based convolution network (GCN) containing three successive deformation blocks, and two graph unpooling layers sandwiched by the three successive deformation blocks. This invention is also related to a system and a method thereof.

Network, system and method for 3D shape generation

For video captioning, with an encoding module and a decoding module, the encoding module comprises a plurality of encoding units each receiving a set of video frames, wherein the sets of video frames received by two neighboring encoding units are in chronological order; and the encoding units each producing a spatially attended feature, so that the plurality of encoding units produce a spatially attended feature sequence; and the decoding module comprises a decoding unit chronologically receiving a temporally attended feature obtained from the spatially attended features sequence. Also disclosed is a method thereof.

System and method for video captioning

A scene recognition method and apparatus are provided. The method includes obtaining multiple local detectors by training a training image set, where one local detector in the multiple local detectors corresponds to one local area of a type of target, and the type of target includes at least two local areas; detecting a to-be-recognized scene by using the multiple local detectors, and acquiring a feature, which is based on a local area of the target, of the to-be-recognized scene; and recognizing the to-be-recognized scene according to the feature, which is based on the local area of the target, of the to-be-recognized scene.

Growing community of inventors

Shanghai, China

Yugang Jiang