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/

Zendy Plus

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

Zendy Tools

Zendy Open

Gradient Inversion (GI) attacks are a ubiquitous threat in Federated Learning(FL) as they exploit gradient leakage to reconstruct supposedly privatetraining data. Common defense mechanisms such as Differential Privacy (DP) orstochastic Privacy Modules (PMs) introduce randomness during gradientcomputation to prevent such attacks. However, we pose that if an attackereffectively mimics a client's stochastic gradient computation, the attacker cancircumvent the defense and reconstruct clients' private training data. Thispaper introduces several targeted GI attacks that leverage this principle tobypass common defense mechanisms. As a result, we demonstrate that noindividual defense provides sufficient privacy protection. To address thisissue, we propose to combine multiple defenses. We conduct an extensiveablation study to evaluate the influence of various combinations of defenses onprivacy protection and model utility. We observe that only the combination ofDP and a stochastic PM was sufficient to decrease the Attack Success Rate (ASR)from 100% to 0%, thus preserving privacy. Moreover, we found that thiscombination of defenses consistently achieves the best trade-off betweenprivacy and model utility.

Combining Stochastic Defenses to Resist Gradient Inversion: An Ablation Study