Software systems are achieving a high level of complexity, thus becoming increasingly hard to manage through a centralised architecture, This is why decentralised software architectures are blooming, \eg multiagent systems, peer-to-peer networks, or sensor networks. Managing decentralised systems is complicated. Therefore, many decentralised systems maintain certain functions centralised, such as security.
A dictionary definition of security is the degree of protection against danger, loss, and criminals. We take an alternative definition: the degree of satisfaction in interactions with others. Achieving security in a decentralised manner requires a different set of techniques than those used in centralised approaches. In this thesis we study enforcement techniques to be applied in a fully decentralised manner. In some cases, decentralised techniques are just generalisations of centralised techniques. Nonetheless, some decentralised techniques are unique to the mechanism used for the interaction process.
By modelling the software systems as multiagent networks, where each agent is connected to those agents it knows, and by defining simple interaction protocols, we have developed new enforcement techniques that can be applied by any agent in the system. The aim of these peer enforcement techniques is to reduce the sanctioned agent's ability to interact, bringing it one step closer to total ostracism. We have also developed sophisticated reputation modelling techniques that are robust against most widespread malicious attacks in order to help enforcing agents decide when to apply the enforcement techniques.
These peer enforcement techniques and reputation mechanisms have been evaluated analytically and experimentally in scenarios ranging from those that are closed and with a shared description of appropriate behaviour, to those that are open and with a subjective description of appropriate behaviour. The analytical results provide information about the limits of these techniques. Whereas the experimental results verify that applying the enforcement techniques has a positive effect in the average satisfaction experienced by the agents in the multiagent network. Furthermore, the experimental results have evaluated the extent of that positive effect and the types of scenarios for which they work best.