RF-based gesture recognition systems outperform computer vision-based systems in terms of user privacy. The integration of Wi-Fi sensing and deep learning has opened new application areas for intelligent multimedia technology. Although promising, existing systems have multiple limitations: (1) they only work well in a fixed domain; (2) when working in a new domain, they require the recollection of a large amount of data. These limitations either lead to a subpar cross-domain performance or require a huge amount of human effort, impeding their widespread adoption in practical scenarios. We propose Wi-AM, a privacy-preserving gesture recognition framework, to address the above limitations. Wi-AM can accurately recognize gestures in a new domain with only one sample. To remove irrelevant disturbances induced by interfering domain factors, we design a multi-domain adversarial scheme to reduce the differences in data distribution between different domains and extract the maximum amount of transferable features related to gestures. Moreover, to quickly adapt to an unseen domain with only a few samples, Wi-AM adopts a meta-learning framework to fine-tune the trained model into a new domain with a one-sample-per-gesture manner while achieving an accurate cross-domain performance. Extensive experiments in a real-world dataset demonstrate that Wi-AM can recognize gestures in an unseen domain with average accuracy of $82.13\%$ and $86.76\%$ for 1 and 3 data samples.

Dataset distillation has emerged as a promising technique to compress large-scale datasets into small synthetic sets, significantly reducing training overhead. However, this paradigm introduces a critical security vulnerability: malicious providers can embed hidden backdoors into the distilled data, causing downstream models to misclassify triggered inputs. Defending against such attacks in the distillation setting presents a unique challenge due to the extreme scarcity of synthetic samples (e.g., 10 images per class). Conventional defenses, which rely on filtering or discarding suspicious samples, fail in this regime as removing high-density samples leads to a catastrophic collapse in model utility. To address this dilemma, we propose SafeDD, a novel defense framework based on a Mask-andRetrain strategy. Our approach formulates an optimization objective to decouple and extract dominant features from synthetic images. We optimize a mask to eliminate these influential features and subsequently retrain the model to evaluate the impact. The core insight leverages the shortcut learning phenomenon: we posit that the dominant features in poisoned samples converge to the backdoor trigger, whereas in benign samples, they correspond to critical semantic content. Consequently, masking these features effectively neutralizes the backdoor in poisoned classes while retaining sufficient semantic information for classification. Extensive experiments demonstrate that SafeDD successfully reduces the attack success rate to near 0% and maintains robust recognition accuracy on the targeted malicious classes.

Zero-shot learning (ZSL) typically leverages semantic knowledge and textual descriptions of classes to forge connections between seen and unseen classes. ZSL can classify new categories of data unseen in the training set. Prior research has focused on aligning image features with their corresponding auxiliary information, overlooking the limitation whereby individual features may not capture the full spectrum of information inherent in the original image. Additionally, there are concerns regarding the bias in the predicted results towards seen classes during the testing phase. In this work, we introduce a novel approach termed Feature Fusion Transformer Network (FFusion), which decomposes visual and semantic features into a multitude of functionally distinct features, fusing them to form new representations that better encapsulate the information content of the original images during training. Furthermore, we implement novel loss functions to balance the model’s focus on unseen and seen classes within the ZSL framework. Experiments show that our model achieves accuracy of 76.8% in the CUB dataset, while reducing the bias between seen and unseen classes to a mere 0.2%.