A critical overview of privacy-preserving approaches for collaborative forecasting

Privacy-Preserving Distributed Learning for Renewable Energy Forecasting

“…suggesting coefficients privacy since the original B is no longer used. However, the limitations identified in a previous work [10] for (5) are valid for (12). That is, a curious agent can obtain both Y and ZQ, and because Y and Z share a large proportion of values, Z can also be recovered.…”

“…To reduce the possibility of such confidentiality breaches, recent work combined distributed ADMM with differential privacy, which consists of adding random noise (with certain statistical properties) to the data itself or coefficients [22], [23]. However, these mechanisms can deteriorate the performance of the model even under moderate privacy guarantees [10].…”

“…Some linear algebra-based protocols exist for secure matricial product, but they were designed for matrices with independent observations and have proven to fail when applied to such matrices as Z and Y(see [10] for a proof). The calculation of MX Ai is described in Algorithm 1.…”

“…Zhang and Wang described a privacy-preserving approach for wind power forecasting with off-site time series, which combined ridge linear quantile regression with alternating direction method of multipliers (ADMM) [9]. However, privacy with ADMM is not always guaranteed since it requires intermediate calculations, allowing the most curious competitors to recover the data at the end of several iterations [10]. Moreover, the central node can also recover the original and private data.…”

“…A literature analysis in [10] of privacy-preserving techniques for VAR has grouped these techniques as (a) data transformation, such as generation of random matrices that pre-or post-multiply the data [13] or using principal component analysis with differential privacy [14] (b) secure multiparty computation, such as linear algebra protocols [15] or homomorphic encryption (encrypting the original data in a way that arithmetic operations in the public space does not compromise the encryption [16]), and (c) decomposition-based methods like the ADMM [17] or the distributed Newton-Raphson method [18]. The main conclusions were that data transformation requires a trade-off between privacy and accuracy, secure multi-party computations either result in computationally demanding techniques or do not fully preserve privacy in VAR models, and that decomposition-based methods rely on iterative processes and after a number of iterations, the agents have enough information to recover private data.…”