Recently, methods for skeleton-based human activity recognition have been shown to be vulnerable to adversarial attacks. However, these attack methods require either the full knowledge of the victim (i.e. white-box attacks), access to training data (i.e. transfer-based attacks) or frequent model queries (i.e. black-box attacks). All their requirements are highly restrictive, raising the question of how detrimental the vulnerability is. In this paper, we show that the vulnerability indeed exists. To this end, we consider a new attack task: the attacker has no access to the victim model or the training data or labels, where we coin the term hard no-box attack. Specifically, we first learn a motion manifold where we define an adversarial loss to compute a new gradient for the attack, named skeleton-motioninformed (SMI) gradient. Our gradient contains information of the motion dynamics, which is different from existing gradient-based attack methods that compute the loss gradient assuming each dimension in the data is independent. The SMI gradient can augment many gradient-based attack methods, leading to a new family of no-box attack methods. Extensive evaluation and comparison show that our method imposes a real threat to existing classifiers. They also show that the SMI gradient improves the transferability and imperceptibility of adversarial samples in both no-box and transfer-based black-box settings.

Abstract In visual object tracking methods, improving both the run time and the accuracy in the face of complex situations has always been an important issue. Many complex tracking algorithms, such as part‐based algorithms, have better accuracy when facing occlusions, but they have much greater computational complexity. In response to the above problems, this paper proposes a selective part‐based correlation filter (SPCF) tracking algorithm with a reinforcement learning to achieve more stable and efficient tracking of targets. First, according to the conditions of the response map of the correlation filter (CF), the entire tracking process is divided into three states: simple environments, complex environments, and harsh environments. Second, this paper uses reinforcement learning to determine the states of frames in different situations to improve the tracking effect of the algorithm. Third, the process of the online selection of states is transformed into a Markov decision process (MDP), where the policy learning of the MDP is achieved by reinforcement learning. Additionally, different strategies are used to track a target in different states: the overall filter is used to increase the speed in simple environments; part‐based filters are used to improve the accuracy in complex environments; and in harsh environments where the target completely disappears, a redetection algorithm is used to find the target when it reappears. Finally, the performance of the tracking algorithm is verified on the VOT2018, OTB‐2015, and LaSOT datasets.

The nonsubsampled Laplacian pyramid (NSLP) is widely used as a common multiresolution decomposition method in various nonsubsampled image transforms. However, the NSLP has a fixed spectrum partition, and thus cannot represent images accurately, and flexibly. We propose a new adaptive arbitrary multiresolution decomposition to solve these problems. First, using the affine characteristics of a pseudopolar Fourier transform (PPFT), we apply a 1-D nonuniform filter bank to the modulated PPFT to obtain a 2-D arbitrary resolution filter bank. This filter surpasses the limitation of fixed spectrum partitioning of the traditional tree structure. We then demonstrate that the proposed method satisfies the compact frame condition, and has translation invariance, and a linear phase. Furthermore, we propose an adaptive spectrum division approach at various scales based on image spectrum information based on a 2-D filter bank of arbitrary multiresolution, so our method can capture important visual information more accurately. Finally, we combine our method with a nonsubsampled directional filter bank of the nonsubsampled contourlet transform to create a new multiscale geometric analysis (MGA) method, and verify that the new method can have perfect reconstruction properties. The new MGA also performs better in image denoising, and recognition experiments than state-of-the-art MGA methods with translation invariance.