Jakub Muszyński

Convergence analysis of evolutionary algorithms in the presence of crash-faults and cheaters

This paper analyzes the fault-tolerance nature of Evolutionary Algorithms (EAs) when executed in a distributed environment subjected to malicious acts. More precisely, the inherent resilience of EAs against two types of failures is considered: (1) crash faults, typically due to resource volatility which lead to data loss and part of the computation loss; (2) cheating faults, a far more complex kind of fault that can be modeled as the alteration of output values produced by some or all tasks of the program being executed. This last type of failure is due to the presence of cheaters on the computing platform. Most often in Global Computing (GC) systems such as BOINC, cheaters are attracted by the various incentives provided to stimulate the volunteers to share their computing resources: cheaters typically seek to obtain rewards with little or no contribution to the system. In this paper, the Algorithm-Based Fault Tolerance (ABFT) aspects of EAs against the above types of faults is characterized. Whereas the inherent resilience of EAs has been previously observed in the literature, for the first time, a formal analysis of the impact of the considered faults over the executed EA including a proof of convergence is proposed in this article. By the variety of problems addressed by EAs, this study will hopefully promote their usage in the future developments around distributed computing platform such as Desktop Grids and Volunteer Computing Systems or Cloud systems where the resources cannot be fully trusted.

Analysis of the Data Flow in the Newscast Protocol for Possible Vulnerabilities

Newscast is a model for information dissemination and membership management in large-scale, agent-based distributed systems. It deploys a simple, peer-to-peer data exchange protocol. The Newscast protocol forms an overlay network and keeps it connected by means of an epidemic algorithm, thus featuring a complex, spatially structured, and dynamically changing environment. It has recently become very popular due to its inherent resilience to node volatility as it exhibits strong self-healing properties. In this paper, we analyze the robustness of the data flow within the Newscast model against a set of vulnerabilities that have not been taken into account in previous analysis. In particular, we perform an attack based on a cache content corruption which is able to defeat the protocol by breaking the network connectivity. Concrete experiments are performed using a framework that implements both the protocol and the corruption model considered in this work.

Reducing Efficiency of Connectivity-Splitting Attack on Newscast via Limited Gossip

Newscast is a Peer-to-Peer, nature-inspired gossip-based data exchange protocol used for information dissemination and membership management in large-scale, agent-based distributed systems. The model follows a probabilistic scheme able to keep a self-organised, small-world equilibrium featuring a complex, spatially structured and dynamically changing environment. Newscast gained popularity since the early 2000 s thanks to its inherent resilience to node volatility as the protocol exhibits strong self-healing properties. However, the original design proved to be surprisingly fragile in a byzantine environment subjected to cheating faults. Indeed, a set of recent studies emphasized the hard-wired vulnerabilities of the protocol, leading to an efficient implementation of a malicious client, where a few naive cheaters are able to break the network connectivity in a very short time. Extending these previous works, we propose in this paper a modification of the seminal protocol with embedded counter-measures, improving the resilience of the scheme against malicious acts without significantly affecting the original Newscast’s properties nor its inherent performance. Concrete experiments were performed to support these claims, using a framework implementing all the solutions discussed in this work.

Distributed Cellular Evolutionary Algorithms in a Byzantine Environment

Distributed parallel computing platforms contribute for a large part to some of the most powerful computers. Such architectures are typically based on accelerators (General Purpose computing on Graphics Processing Units, Many Integrated Cores e.g. Xeon Phi co-processors) and/or a large number of interconnected computing nodes. Obviously, they raise new challenges, typically in terms of scalability, robustness, adaptability and security. At the advent of the quest for Ultra scale Computing Systems, this paper addresses the issue of fault tolerance toward Byzantine failures overs such platforms. Indeed, the inherent unpredictable nature of these errors render their detection, not speaking of their correction, hard or even impossible to perform at large-scale. At this level, Algorithm-Based Fault Tolerance (ABFT) techniques where the fault tolerance scheme is tailored to the algorithm performed, seems the most promising approaches to deal with such failures. In this context, Evolutionary Algorithms (EAs), especially panmictic global parallel EAs, exhibit a remarkable resilience against byzantine failures modeled as cheating faults as demonstrated either empirically or theoretically in previous studies. In this paper, we extend this analysis to the case of distributed EAs based on the cellular model leading to distributed Cellular Evolutionary Algorithms (dCEAs). Our empirical study over a set or reference optimization problem confirm the ABFT nature of dCEAs. To our knowledge, this is the first study of dCEAs under the perspective of cheating issues and crash faults in a domain of distributed computations, thus opening new insights and perspectives for the design of competitive ultra-scale system based on evolutionary programming models.

Exploiting the Hard-Wired Vulnerabilities of Newscast via Connectivity-Splitting Attack

Newscast is a model for information dissemination and membership management in large-scale, agent-based distributed systems. It deploys a simple, peer-to-peer data exchange protocol. The Newscast protocol forms an overlay network and keeps it connected by means of an epidemic algorithm, thus featuring a complex, spatially structured, and dynamically changing environment. It has recently become very popular due to its inherent resilience to node volatility as it exhibits strong self-healing properties. In this paper, we analyze the robustness of the Newscast model when executed in a distributed environment subjected to malicious acts. More precisely, we evaluate the resilience of Newscast against cheating faults and demonstrate that even a few naive cheaters are able to defeat the protocol by breaking the network connectivity. Concrete experiments are performed using a framework that implements both the protocol and the cheating model considered in this work.

Expected running time of parallel evolutionary algorithms on unimodal pseudo-boolean functions over small-world networks

This paper proposes a theoretical and experimental analysis of the expected running time for an elitist parallel Evolutionary Algorithm (pEA) based on an island model executed over small-world networks. Our study assumes the resolution of optimization problems based on unimodal pseudo-boolean funtions. In particular, for such function with d values, we improve the previous asymptotic upper bound for the expected parallel running time from O(d√n) to O(d log n). This study is a first step towards the analysis of influence of more complex network topologies (like random graphs created by P2P networks) on the runtime of pEAs. A concrete implementation of the analysed algorithm have been performed on top of the ParadisEO framework and run on the HPC platform of the University of Luxembourg (UL). Our experiments confirm the expected speedup demonstrated in this article and prove the benefit that pEA can gain from a small-world network topology.