English (United Kingdom)

https://curated-unify.zendy.io/wp-json/zendy-region/v1/featured_content/oa?rat=en

https://curated-unify.zendy.io/wp-json/zendy-region/v1/highlighted_journal/

Global: Zendy Plus annual

Presents the access of premium content as premium feature

Premium Content

Presents the keyphrase highlighting as premium feature

Keyphrase Highlighting

Presents the summarisation as premium feature

Summarisation

Insights

Presents the pdf analysis as premium feature

PDF Analysis

Presents the zaia usage as premium feature

ZAIA

Global: Zendy Plus monthly

Global: Zendy Tools annual

Global: Zendy Tools monthly

Global: Zendy Free Plan

Recent deep neural networks (DNNs) have come to rely on vast amounts oftraining data, providing an opportunity for malicious attackers to exploit andcontaminate the data to carry out backdoor attacks. These attacks significantlyundermine the reliability of DNNs. However, existing backdoor attack methodsmake unrealistic assumptions, assuming that all training data comes from asingle source and that attackers have full access to the training data. In thispaper, we address this limitation by introducing a more realistic attackscenario where victims collect data from multiple sources, and attackers cannotaccess the complete training data. We refer to this scenario asdata-constrained backdoor attacks. In such cases, previous attack methodssuffer from severe efficiency degradation due to the entanglement betweenbenign and poisoning features during the backdoor injection process.  To tackle this problem, we propose a novel approach that leverages thepre-trained Contrastive Language-Image Pre-Training (CLIP) model. We introducethree CLIP-based technologies from two distinct streams: Clean FeatureSuppression, which aims to suppress the influence of clean features to enhancethe prominence of poisoning features, and Poisoning Feature Augmentation, whichfocuses on augmenting the presence and impact of poisoning features toeffectively manipulate the model's behavior.  To evaluate the effectiveness, harmlessness to benign accuracy, andstealthiness of our method, we conduct extensive experiments on 3 targetmodels, 3 datasets, and over 15 different settings. The results demonstrateremarkable improvements, with some settings achieving over 100% improvementcompared to existing attacks in data-constrained scenarios. Our researchcontributes to addressing the limitations of existing methods and provides apractical and effective solution for data-constrained backdoor attacks.

Efficient Backdoor Attacks for Deep Neural Networks in Real-world Scenarios