The research on image-classification-adversarial attacks is crucial in the realm of artificial intelligence (AI) security. Most of the image-classification-adversarial attack methods are for white-box settings, demanding target model gradients and network architectures, which is less practical when facing real-world cases. However, black-box adversarial attacks immune to the above limitations and reinforcement learning (RL) seem to be a feasible solution to explore an optimized evasion policy. Unfortunately, existing RL-based works perform worse than expected in the attack success rate. In light of these challenges, we propose an ensemble-learning-based adversarial attack (ELAA) targeting image-classification models which aggregate and optimize multiple reinforcement learning (RL) base learners, which further reveals the vulnerabilities of learning-based image-classification models. Experimental results show that the attack success rate for the ensemble model is about 35% higher than for a single model. The attack success rate of ELAA is 15% higher than those of the baseline methods.
22 February 2024 in Expert Systems with Applications
Heterogeneous graph representation learning is to learn effective representations for nodes or (sub)graphs, which preserve node attributes and structural information. However, it is challenging to design a representation learning method for heterogeneous information networks (HINs) due to their diversity. Most of the existing HIN-oriented learning methods define a series of meta-paths. Then, they aggregate the representations learned from different meta-paths in the same hidden space. These methods do not consider semantic differences of different meta-paths, which leads to semantic confusion. And further affects the effectiveness of the learned representation. Given these issues, we introduce a Semantic Aware HIN Representation learning Network (SAHRN), which takes into account the semantics of different meta-paths. We mitigate the problem of semantic confusion by projecting nodes’ features into different hidden spaces separately according to different meta-paths. To further expand the scope of aggregation and enrich the aggregated information, we also design various variants of our model by adding layer aggregation. Extensive experiments on three standard HIN datasets show that SAHRN achieves consistent improvements compared to state-of-the-art graph representation learning methods. The experiments and analyses on each component of the model show the effectiveness of the proposed method. The source code is available on https://github.com/pingpingand/SAHRN.