The right to privacy, enshrined in various human rights declarations, faces new challenges in the age of artificial intelligence (AI). This paper explores the concept of the Right to be Forgotten (RTBF) within AI systems, contrasting it with traditional data erasure methods. We introduce Forgotten by Design, a proactive approach to privacy preservation that integrates instance-specific obfuscation techniques during the AI model training process. Unlike machine unlearning, which modifies models post-training, our method prevents sensitive data from being embedded in the first place. Using the LIRA membership inference attack, we identify vulnerable data points and propose defenses that combine additive gradient noise and weighting schemes. Our experiments on the CIFAR-10 dataset demonstrate that our techniques reduce privacy risks by at least an order of magnitude while maintaining model accuracy (at 95% significance). Additionally, we present visualization methods for the privacy-utility trade-off, providing a clear framework for balancing privacy risk and model accuracy. This work contributes to the development of privacy-preserving AI systems that align with human cognitive processes of motivated forgetting, offering a robust framework for safeguarding sensitive information and ensuring compliance with privacy regulations.

We implement and test the recently proposed Inhibitor gate for Recurrent Neural Networks (RNNs) that is both efficient and accurate under Homomorphic Encryption. Gated RNNs such as LSTM and GRU have applications in numerous real-world use cases for sequential data, such as time-series analysis and natural language processing, due to their ability to capture long-term dependencies. Conventionally, the update based on current inputs and the previous state history is each multiplied with the dynamic gate values and combined to compute the next state. However, as it is a multiplication between two variables that depend on the input, it can be com-putationally expensive, especially under homomorphic en-cryption, where it involves a multiplication between two cipher text variables. Therefore, the novel gating mechanism replaces the multiplication and sigmoid function of the conventional RNN with addition and ReLU activation. Numerical experiments on three synthetic benchmark tests demonstrate that our algorithm outperforms a conventional gated RNN showing at least an order of magnitude faster execution already at 4-bit quantization. At a fixed computational budget , we get 7-bit precision with the novel gate that executes faster than the 4-bit conventional gate. The gating mechanism employed in this paper may enable privacy-preserving AI applications based on gated RNNs operating under homomorphic encryption by avoiding the multiplication of encrypted variables.

The optimal utilisation of distribution grids requires the proactive management of volatilities caused by mixed energy resources installed into different grid levels, such as buildings, energy communities (ECs) and substations. In this context, proactive control based on predictions for energy demand and generation is applied. The mitigation of conflicts between the stakeholders' objectives is the main challenge for the control of centralized and distributed energy resources. In this paper, a bi-level approach is proposed for the control of stationary battery energy storage systems (SBES) supporting the local distribution system operator (DSO) at the transformer level, as well as distributed energy resources (DERs) operated by end customers, i.e., EC-members. Model predictive control (MPC)- based and hybrid approaches merging rule- and MPC-based control schemes are evaluated. Simulation studies based on a typical European low voltage (LV) feeder topology yield the performance assessment in terms of technical and economic criteria. The results show an advantage of hybrid approaches with respect to the DSO's cost savings from peak shaving. From the EC's perspective, both hybrid and MPC-based schemes can achieve effective cost savings from proactive energy management.

To enhance the computational efficiency of quantized Transformers, we replace the dot-product and Softmax-based attention with an alternative mechanism involving addition and ReLU activation only. This side-steps the expansion to double precision often required by matrix multiplication and avoids costly Softmax evaluations but maintains much of the core functionality of conventional dot-product attention. It can enable more efficient execution and support larger quantized Transformer models on resource-constrained hardware or alternative arithmetic systems like homomorphic encryption. Training experiments on four common benchmark tasks show test set prediction scores comparable to those of conventional Transformers with dot-product attention. Our scaling experiments also suggest significant computational savings, both in plaintext and under encryption. The ReLU and addition-based attention mechanism introduced in this paper may enable privacy-preserving AI applications operating under homomorphic encryption by avoiding the costly multiplication of encrypted variables.

As the demand for data privacy and low latency grows, edge computation carried out at edge data center nodes is believed to become increasingly important for future telecom applications. Providers must consider various factors, including power consumption, thermal dynamics, and the ability to maintain high-quality service, in addition to deployment and service orchestration. This paper presents a detailed description of two different prototype edge data centers designed to investigate the power performance and thermal dynamics of edge nodes under various applied services. The prototypes were developed and tested at the RISE ICE Datacenter research facility. We present the results of power flow experiments in which input current from the grid was limited while the computational load was maintained using the energy stored in batteries. We further discuss implications for placing edge data center nodes in locations with temporal power constraints and opportunities for participation in support services at the grid level.

A recent report by the Swedish Authority for Privacy Protection (IMY) evaluates the potential of jointly training and exchangingmachine learningmodels between two healthcare providers. In relation to the privacy problems identified therein, this article explores the trade-off between utility and privacy when using privacyenhancing technologies (PETs) in combination with federated learning. Results are reported from numerical experiments with standard text-book machine learning models under both differential privacy (DP) and FullyHomomorphic Encryption (FHE). The results indicate that FHE is a promising approach for privacy-preserving federated learning, with the CKKS scheme being more favorable in terms of computational performance due to its support of SIMD operations and compact representation of encrypted vectors. The results for DP are more inconclusive. The article briefly discusses the current regulatory context and aspects that lawmakers may consider to enable an AI leap in Swedish healthcare while maintaining data protection.

This article investigates the problem of where to place the computation workload in an edge-cloud network topology considering the trade-off between the location-specific cost of computation and data communication. For this purpose, a Monte Carlo simulation model is defined that accounts for different workload types, their distribution across time and location, as well as correlation structure. Results confirm and quantify the intuition that optimization can be achieved by distributing a part of cloud computation to make efficient use of resources in an edge data center network, with operational energy savings of 4–6% and up to 50% reduction in its claim for cloud capacity.

Adequate privacy protection is crucial for implementing modern AI algorithms in medicine. With Fully Homomorphic Encryption (FHE), a party without access to the secret key can perform calculations and advanced analytics on encrypted data without taking part of either the input data or the results. FHE can therefore work as an enabler for situations where computations are carried out by parties that are denied plain text access to sensitive data. It is a scenario often found with digital services that process personal health-related data or medical data originating from a healthcare provider, for example, when the service is delivered by a third-party service provider located in the cloud. There are practical challenges to be aware of when working with FHE. The current work aims to improve accessibility and reduce barriers to entry by providing code examples and recommendations to aid developers working with health data in developing FHE-based applications. HEIDA is available on the GitHub repository: https://github.com/rickardbrannvall/HEIDA.

The design of intelligent algorithms used for device monitoring and control can be costly and is an investment that must be protected against reverse engineering by competitors. An algorithm can be safeguarded by running remotely from the cloud instead of locally on the equipment hardware. However, such a setup requires that sensitive data is sent from the device to the cloud. Fully Homomorphic Encryption (FHE) is an emerging technology that offers a solution to this problem since it enables computation on encrypted data. A cloud service using FHE can protect its proprietary algorithms while simultaneously offering customer data confidentiality. The computational overhead for the technology is, however, still very high. This work reports on a practical investigation of using FHE for data center remote control problems: What applications are feasible today? And at what cost?

This study investigates the use of transfer learning and modular design for adapting a pretrained model to optimize energy efficiency and heat reuse in edge data centers while meeting local conditions, such as alternative heat management and hardware configurations. A Physics-Informed Data-Driven Recurrent Neural Network (PIDD RNN) is trained on a small scale-model experiment of a six-server data center to control cooling fans and maintain the exhaust chamber temperature within safe limits. The model features a hierarchical regularizing structure that reduces the degrees of freedom by connecting parameters for related modules in the system. With a RMSE value of 1.69, the PIDD RNN outperforms both a conventional RNN (RMSE: 3.18), and a State Space Model (RMSE: 2.66). We investigate how this design facilitates transfer learning when the model is fine-tuned over a few epochs to small dataset from a second set-up with a server located in a wind tunnel. The transferred model outperforms a model trained from scratch over hundreds of epochs.