Considering data safety in more and more applied peer-to-peer networks, such as wireless sensor networks, has become the focus of information fusion, this paper proposes the problem of credible evidence fusion (CEF) in a distributed system with privacy-preserving, where agent’s raw evidence is shared only with authenticated neighbors while access or inference by non-neighbors is prevented. This privacy-preserving in a distributed style brings out two challenges never met in centralized CEF. One is about credibility calculation consensus (CCC), where the local evidence difference measure matrix (EDMM) faces elements missing for evidence unavailability to non-neighbors. The other is about privacy-preserving fusion consensus (PFC), because fusion based on raw evidence may lead to a counter-intuitive result. In CCC, missing elements in EDMM are recovered via low-rank matrix completion, and credibility consistency is further guaranteed via diffusion gradient descent. In PFC, a new measure named privacy iterative compensation term (PICT) is constructed based on credibility and discounted evidence so that raw evidence is impossibly inferred. Furthermore, the linear average consensus based on PICT is presented and proved to be converging to the centralized CEF if credibility is accurately given. The computational complexity is proportional to the cube of the number of agents and the square of the number of elements in the power set of the framework of discrimination. Simulation results demonstrate that the proposed method achieves credibility errors below 0.1, leading to fusion results and classification accuracy approaching those of centralized CEF, with a 30% improvement compared to RANSAC-based and COF-based methods.

中文翻译：

---

分布式可信证据融合与隐私保护


考虑到无线传感器网络等越来越多应用的点对点网络中的数据安全已成为信息融合的焦点，本文提出了隐私保护的分布式系统中的可信证据融合（CEF）问题，其中代理的原始证据仅与经过身份验证的邻居共享，同时防止非邻居访问或推断。这种分布式风格的隐私保护带来了集中式 CEF 从未遇到过的两个挑战。一是关于可信度计算共识（CCC），其中局部证据差异度量矩阵（EDMM）面临因非邻居无法获得证据而缺少元素的问题。另一个是关于隐私保护融合共识（PFC），因为基于原始证据的融合可能会导致反直觉的结果。在CCC中，EDMM中缺失的元素通过低秩矩阵补全来恢复，并通过扩散梯度下降进一步保证可信度一致性。在PFC中，基于可信度和折扣证据构建了一种名为隐私迭代补偿项（PICT）的新措施，使得原始证据不可能被推断出来。此外，提出了基于 PICT 的线性平均共识，并证明如果准确给出可信度，则可以收敛到集中式 CEF。计算复杂度与智能体数量的立方和判别框架的幂集中元素数量的平方成正比。仿真结果表明，该方法的可信度误差低于0.1，融合结果和分类精度接近集中式CEF，与基于RANSAC和基于COF的方法相比提高了30%。