Infrastructure as a Service and memory forensics are two subjects which have recently gained increasing amounts of attention. Combining these topics poses new challenges when performing forensic investigations. Forensics targeting virtual machines in a cloud environment is problematic since the devices are virtual, and memory forensics are a newer branch of forensics which is hard to perform and is not well documented. It is, however an area of utmost importance since virtual machines may be targets of, or participate in suspicious activity to the same extent as physical machines. Should such activity require an investigation to be conducted, some data which could be used as evidence may only be found in the primary memory. This thesis aims to further examine memory forensics in cloud environments and expand the academic field of these subjects and help cloud hosting organisations. The objective of this thesis was to study if Virtual Machine Introspection is a valid technique to acquire forensic evidence from the virtual primary memory of a virtual machine. Virtual Machine Introspection is a method of monitoring and analysing a guest via the hypervisor. In order to verify whether Virtual Machine Introspection is a valid forensic technique, the first task was to attempt extracting data from the primary memory which had been acquired using Virtual Machine Introspection. Once extracted, the integrity of the data had to be authenticated. This was done by comparing a hash sum of a file located on a guest with a hash sum of the extracted data. The experiment showed that the two hashes were an exact match. Next, the solidity of the extracted data was tested by changing the memory of a guest while acquiring the memory via Virtual Machine Introspection. This showed that the solidity is heavily compromised because memory acquisition process used was too slow. The final task was to compare Virtual Machine Introspection to acquiring the physical memory of the host. By setting up two virtual machines and examining the primary memory, data from both machines was found where as Virtual Machine Introspection only targets one machine, providing an advantage regarding privacy.

L’introspection de machine virtuelle (VM) consiste à superviser les états et les activités de celles-ci depuis la couche de virtualisation, tirant ainsi avantage de son emplacement qui offre à la fois une bonne visibilité des états et des activités des VMs ainsi qu’une bonne isolation de ces dernières. Cependant, les états et les activités des VMs à superviser sont vus par la couche de virtualisation comme une suite binaire de bits et d’octets en plus des états des ressources virtuelles. L’écart entre la vue brute disponible à la couche de virtualisation et celle nécessaire pour la supervision de sécurité des VMs constitue un challenge pour l’introspection appelé « le fossé sémantique ». Pour obtenir des informations sémantiques sur les états et les activités des VMs à fin de superviser leur sécurité, nous présentons dans cette thèse un ensemble de techniques basé sur l’analyse binaire et la réutilisation du code binaire du noyau d’une VM. Ces techniques permettent d’identifier les adresses et les noms de la plupart des fonctions noyau d’une VM puis de les instrumenter (intercepter, appeler et analyser) pour franchir le fossé sémantique de manière automatique et efficiente même dans les cas des optimisations du compilateur et de la randomisation de l’emplacement du code noyau dans la mémoire de la VM. / Virtual Machine Introspection (VMI) consists inmonitoring VMs security from the hypervisor layer which offers thanks to its location a strong visibility on their activities in addition to a strong isolation from them. However, hypervisor view of VMs is just raw bits and bytes in addition to hardware states. The semantic difference between this raw view and the one needed for VM security monitoring presents a significant challenge for VMI called “the semantic gap”. In order to obtain semantic information about VM states and activities for monitoring their security from the hypervisor layer, we present in this thesis a set of techniques based on analysis and reuse of VM kernel binary code. These techniques enable to identify addresses and names of most VM kernel functions then instrument (call, intercept and analyze) them to automatically bridge the semantic gap regardless of challenges presented by compiler optimizations and kernel base address randomization.