Software security relevance for AI/ML solutions in networking. With softwarization of networks, there emerges tight coupling between software security and network security regarding AI/ML schemes. Leveraging AI/ML solutions for autonomic real time management and optimization of the next generation of telecom network is a technical domain per-se. While there is a stock phrase stating that leveraging AI/ML is an unavoidable for the management of 5G networks, there is the same appreciation for the security risks this situation poses. Moreover, if AI/ML driven control is compromised, that may lead to large-scale impact due to automation and closed-loop control. As all pre-cited techniques, there is no absolute defence and all defences shall be studied with their mitigation efficiency, possible accuracy decrease on clean data and their associated performance impact. Finally, the MTD principle applied on AI frustrates the carving of attacks targeted for fixed targets (i.e., static algorithm, static model, static training data), through dynamicity and variability. This research is certainly at its initial stage and as stated by the authors will be discussed on the heavy workflow and performance impact issues. The technical survey made above, in a general perspective, in the context of applying AI/ML for 5G and Beyond networking, leads us the following conclusions: x In this domain of Adversarial Machine Learning (AML), both attackers and defenders develop data-based techniques derived from data analytics and data statistics with the idea to carve attack vectors with high hit rate and equivalent detection ability on the other side. Both sides reside in some kind of uncertainty. Every technique which hardens the life of the attacker and ĚƌŽƉ ƚŚĞ ĂƚƚĂĐŬ ĞĨĨŝĐŝĞŶĐLJ ǁŝůů ďĞ ǁĞůĐŽŵĞ merits, overhead and setup issues. x Before taking a closer look on the most promising software security techniques and associated best use for AI/ML pipeline security, it is worth restating the large functional scope brought by software security measures. Various complementary techniques can be applied on software for its security. From vulnerability remediation to variant singularization, execution control and run-time monitoring, the scope of possible hardening techniques is large and largely unemployed as of today.
