Cloud providers leverage live migration of virtual machines to reduce energy consumption and allocate resources efficiently in data centers. Each decision depends on three questions: when move a machine, which machine where it? Dynamic, uncertain, heterogeneous workloads running make such decisions difficult. Knowledge-based heuristics-based algorithms are commonly used tackle this problem. algorithms, as MaxWeight scheduling dependent the specifics dynamics targeted architectures...

UDO is a versatile tool for offline tuning of database systems specific workloads. can consider variety choices, reaching from picking transaction code variants over index selections up to system parameter tuning. uses reinforcement learning converge near-optimal configurations, creating and evaluating different configurations via actual query executions (instead relying on simplifying cost models). To cater types, distinguishes heavy parameters (which are expensive change, e.g. physical...

As a technology ML is oblivious to societal good or bad, and thus, the field of fair machine learning has stepped up propose multiple mathematical definitions, algorithms, systems ensure different notions fairness in applications. Given multitude propositions, it become imperative formally verify metrics satisfied by algorithms on datasets. In this paper, we stochastic satisfiability (SSAT) framework, Justicia, that verifies measures supervised with respect underlying data distribution. We...

Driven by the unprecedented availability of data, machine learning has become a pervasive and transformative technology across industry science. Its importance to marine science been codified as one goal UN Ocean Decade. While increasing amounts of, for example, acoustic data are collected research monitoring purposes, methods can achieve automatic processing analysis they require large training datasets annotated or labelled experts. Consequently, addressing relative scarcity is, besides...

In recent years, machine learning (ML) algorithms have been deployed in safety-critical and high-stake decision-making, where the fairness of is paramount importance. Fairness ML centers on detecting bias towards certain demographic populations induced by an classifier proposes algorithmic solutions to mitigate with respect different definitions. To this end, several verifiers proposed that compute prediction classifier—essentially beyond a finite dataset—given probability distribution input...

Fairness in machine learning has attained significant focus due to the widespread application high-stake decision-making tasks. Unregulated classifiers can exhibit bias towards certain demographic groups data, thus quantification and mitigation of classifier is a central concern fairness learning. In this paper, we aim quantify influence different features dataset on classifier. To do this, introduce Influence Function (FIF). This function breaks down into its components among individual...

Based on differential privacy (DP) framework, we introduce and unify definitions for the multi-armed bandit algorithms. We represent framework with a unified graphical model use it to connect definitions. derive contrast lower bounds regret of algorithms satisfying these leverage proving technique achieve all bounds. show that them, learner's is increased by multiplicative factor dependent level $\epsilon$. observe dependency weaker when do not require local rewards.

We propose a reinforcement learning algorithm, Megh, for live migration of virtual machines that simultaneously reduces the cost energy consumption and enhances performance. Megh learns uncertain dynamics workloads as-it-goes. uses dimensionality reduction scheme to projectthe combinatorially explosive state-action space polynomial dimensional space. These schemes enable be scalable work in real-time. experimentally validate is more cost-effective time-efficient than MadVM MMT algorithms.

UDO is a versatile tool for offline tuning of database systems specific workloads. can consider variety choices, reaching from picking transaction code variants over index selections up to system parameter tuning. uses reinforcement learning converge near-optimal configurations, creating and evaluating different configurations via actual query executions (instead relying on simplifying cost models). To cater types, distinguishes heavy parameters (which are expensive change, e.g. physical...

Abstract The calibration of noise for a privacy-preserving mechanism depends on the sensitivity query and prescribed privacy level. A data steward must make non-trivial choice level that balances requirements users monetary constraints business entity. Firstly, we analyse roles sources randomness, namely explicit randomness induced by distribution implicit data-generation distribution, are involved in design mechanism. finer analysis enables us to provide stronger guarantees with...

Federated Learning (FL) is a collaborative machine learning (ML) approach, where multiple clients participate in training an ML model without exposing the private data. Fair and accurate assessment of client contributions important problem FL to facilitate incentive allocation encouraging diverse unified training. Existing methods for assessing contribution adopts co-operative game-theoretic concepts, such as Shapley values, but under simplified assumptions. In this paper, we propose...

With the pervasive deployment of Machine Learning (ML) models in real-world applications, verifying and auditing properties ML have become a central concern. In this work, we focus on three properties: robustness, individual fairness, group fairness. We discuss two approaches for model estimation with without reconstruction target under audit. Though first approach is studied literature, second remains unexplored. For purpose, develop new framework that quantifies different terms Fourier...

We study the per-datum Membership Inference Attacks (MIAs), where an attacker aims to infer whether a fixed target datum has been included in input dataset of algorithm and thus, violates privacy. First, we define membership leakage as advantage optimal adversary targeting identify it. Then, quantify for empirical mean, show that it depends on Mahalanobis distance between data-generating distribution. further assess effect two privacy defences, i.e. adding Gaussian noise sub-sampling....

Bandits serve as the theoretical foundation of sequential learning and an algorithmic modern recommender systems. However, systems often rely on user-sensitive data, making privacy a critical concern. This paper contributes to understanding Differential Privacy (DP) in bandits with trusted centralised decision-maker, especially implications ensuring zero Concentrated (zCDP). First, we formalise compare different adaptations DP bandits, depending considered input interaction protocol. Then,...

High dimensional sparse linear bandits serve as an efficient model for sequential decision-making problems (e.g. personalized medicine), where high features genomic data) on the users are available, but only a small subset of them relevant. Motivated by data privacy concerns in these applications, we study joint differentially private bandits, both rewards and contexts considered data. First, to quantify cost privacy, derive lower bound regret achievable this setting. To further address...