While machine learning models have demonstrated powerful capabilities in widespread applications, they also reveal serious vulnerabilities in both adversarial robustness and privacy preservation. Although existing studies have attempted to build unified frameworks to simultaneously address these two major challenges, their privacy protections often rely on empirical defenses tailored to specific attacks—such as attribute inference—and lack formal, universal security guarantees. To fill this gap, we propose PIRL (Privacy-Preserving Information-theoretic Robust Learning), an information-theoretic representation learning framework designed to achieve jointly provable robustness and formal differential privacy (DP) guarantees. Our core idea is to learn a stochastic representation rigorously constrained by (ϵ,δ)-differential privacy through a unified optimization objective. This representation is forced to "forget" private information while being guided to preserve utility information essential for downstream tasks and to resist adversarial perturbations. We theoretically derive and characterize the intrinsic trade-offs among the model’s task utility, certified robustness, and differential privacy budget (ϵ,δ). Furthermore, we design a practical end-to-end training algorithm and conduct extensive experiments on multiple benchmark datasets. Experimental results show that, compared to existing methods, our framework achieves a superior Pareto frontier balancing these three core metrics.

Local differential privacy (LDP) techniques obviate the need for trust in the data collector, as they provide robust privacy guarantees against untrusted data managers while simultaneously preserving the accuracy of statistical information derived from the privatized data. As a result, these methods have garnered considerable interest and research efforts. In particular,$(\varepsilon,\delta)$-LDP schemes have been utilized across a range of statistical tasks. Nonetheless, existing$(\varepsilon,\delta)$-LDP mechanisms for mean estimation suffer from challenges such as elevated estimation errors and diminished data utility. To address this problem, we propose two novel$(\varepsilon,\delta)$-LDP algorithms for mean estimation. Specifically, we design a one-dimensional piecewise mean estimation algorithm, which perturbs the input data into intervals, thereby reducing noise addition and enhancing both accuracy and efficiency. Building on this foundation, we extend our approach to multi-dimensional data, resulting in a multi-dimensional piecewise mean estimation algorithm. Furthermore, we conduct a theoretical analysis to derive both the variance and error bounds for the proposed algorithms. Extensive experiments conducted on real datasets demonstrate the high practicality of our algorithms for data statistical tasks, showing significant improvements in data utility.

Abstract As an important approach to overcome data silos and privacy concerns in deep learning, federated learning, which can jointly train the global model and keep data local, has shown remarkable performance in a range of industrial applications. However, federated learning still suffers from the problem that shared gradients may be subject to tampering, inference functions, and falsification. To address this issue, we propose a verifiable federated learning framework to deal with malicious aggregators. Initially, we propose a reputation calculation mechanism to solve the problem of selecting a reliable aggregator based on a multiweight subjective logic model. Furthermore, we design a verifiable federated learning scheme to ensure data confidentiality, integrity, and verifiability, as well as support the client's dynamic withdrawal. Security analyses indicate that our framework is secure against malicious adversaries. Furthermore, experimental results on real datasets show that our verifiable federated learning has high accuracy and feasible efficiency.

Differential privacy mechanism can maintain privacy-utility monotonicity. Thus, differential privacy mechanism does not obtain privacy-utility balance for numerical data. To this end, we provide privacy-utility balance of differential privacy mechanism with the collaborative perspective in this paper. First, we constructed the collaborative model achieving privacy-utility balance of differential privacy mechanism. Second, we presented the collaborative algorithm of differential privacy mechanism under our collaborative model. Third, our theoretical analysis showed that the collaborative algorithm of differential privacy mechanism could keep privacy-utility balance. Finally, our experimental results demonstrated that the collaborative differential privacy mechanism can maintain privacy-utility balance. Thus, we provide a new collaborative model to solve the privacy-utility balance problem of differential privacy mechanism. Our collaborative algorithm is easy to apply to query processing of numerical data.