Cryptographic circuits, because they contain confidential informations, are subject to fraud from malicious users, commonly known as attacks. Several attacks have been published and analysed. One of the most effective attack, called Differential Fault Analysis (DFA), uses some fault, voluntary injected by the attacker during the computations, for example with a laser. However, fault models used by these attacks can be restrictive and determine the effectiveness of the attack. Thus, it is important to know which fault model is useful or feasible according to the targeted device or injection means (in our case the laser).A first study about the injected fault types (Bit-set, Bit-reset or Bit-flip) on SRAM memory cells highlighted the strong data dependency of the injected faults and the irrelevance of the Bit-flip fault type. This last result allows to mount Safe Error attacks and creates a real security issue. These results were obtain thanks to sensitivity laser map performed on an isolated SRAM cell and on an 8-bits micro-controller RAM memory. To confirm these experimental results, SPICE simulations have been made with a model developed in the department. This model takes into account the topology of the target.Tests were then carried out on an ASIC implementing the AES algorithm. The fault analysis showed the presence of the three types of faults but also a low injection rates. In contrast, the error repeatability was particularly high. This allowed us to simplify an existing attack and to obtain an attack more effective than conventional attacks, requiring fewer faulted cipher text and reducing the complexity of the analysis to find the secret key. Finally, an assessment of the countermeasure of this circuit showed their ineffectiveness with respect to fault laser attacks. Areas for improvement were then proposed.
