In this thesis we have studied different deobfuscation approaches toward a static evaluation of obfuscation transformations.We mainly focused on static semantic reasoning, combining it with well known techniques from other research areas such as binary diffing and machine learning.We studied and developed several deobfuscation frameworks, one for each of the followings approaches; simplifying the obfuscated code, removing the obfuscation transformations or gathering informations about the protections applied.This thesis contributions are made to answer the following questions:1. How can we contribute to existing generic deobfuscation methodologies?2. How can we use machine learning techniques for the purpose of removing widely used obfuscation transformations?3. How can we help reverse-engineers select the adequate deobfuscation analyses?In order to answer the first question, we transposed semantic-based binary diffing techniques for the purpose of statically simplifying obfuscated binaries.We developed our methodology, called DoSE, as an IDA Pro plug-in and for three major applications.We evaluated each applications against real-world and well-known malwares such as Cryptowall and Flame.Our evaluations underlines the efficiency of DoSE for each applications, with up to 63% of control-flow graph reduction or 1954 cloned functions detected on Flame.We demonstrated that DoSE can be efficiently extended to the detection of two-way opaque predicates, which until then were not detected by any known technique.Therefore, this contribution paves the way for combining semantic equivalence methodologies with existing generic deobfuscation techniques, in order to improve their efficiency and scalability.Our second contribution is, to the best of our knowledge, the first deobfuscation technique based on machine learning for the purpose of removing obfuscation transformations.By introducing the different constructions of opaque predicates and the limitations from dynamic symbolic execution techniques and SMT solvers,we underlined the importance of studying other alternatives for generic evaluations of these transformations.We developed an IDA Pro plug-in, a new approach that bridges a thresholded static symbolic execution with machine learning classification to evaluate both the stealth and resilience of invariant opaque predicates constructions.The use of static symbolic execution allows us to have a better code coverage and scalability, which combined with a machine learning model, permits a generic approach by discarding the use of SMT solvers.Our studies illustrate that our choices conduct towards the implementation of an efficient and accurate evaluation framework against state of the art obfuscators.We created two models for the evaluation of stealth and resiliency of state-of-the-art opaque predicates constructions, with results up to 99% for detection and 95% for deobfuscation.Moreover, we extended our work to a deobfuscation plug-in and compared our results to other tools, showing the efficiency of machine learning for the deobfuscation of most invariant opaque predicates constructions.The third contribution of this thesis aims at improving existing metadata recovery attack. Existing technique syntax-oriented and thus suffers from several limitations.We presented the efficiency of semantic reasoning combined with advanced machine learning techniques.This combination is motivated by the construction of a fine-grained detection framework of obfuscation transformations and constructions.By extending our approach to multi-label and multi-output classification, we enhanced metadata recovery attacks to the detection of multiple layers of obfuscation transformations, as well as their constructions.Our results are promising, with overall accuracies up to 91% for the transformations and 100% for the constructions.