An increasing number of recent experimental works have been demonstrating the supposedly secure channels in the Internet are prone to privacy breaking under many respects, due to traffic features leaking information on the user activity and traffic content. As a matter of example, traffic flow classification at application level, web page identification, language/phrase detection in VoIP communications have all been successfully demonstrated against encrypted channels. In this thesis I aim at understanding if and how complex it is to obfuscate the information leaked by traffic features, namely packet lengths, direction, times. I define a security model that points out what the ideal target of masking is, and then define the optimized and practically implementable masking algorithms, yielding a trade-off between privacy and overhead/complexity of the masking algorithm. Numerical results are based on measured Internet traffic traces. Major findings are that: i) optimized full masking achieves similar overhead values with padding only and in case fragmentation is allowed; ii) if practical realizability is accounted for, optimized statistical masking algorithms attain only moderately better overhead than simple fixed pattern masking algorithms, while still leaking correlation information that can be exploited by the adversary.