Adversarial Machine Learning in Smart Energy Systems

Abstract

Smart Energy Systems represent a radical shift in the approach to energy generation and demand, driven by decentralisation of the energy system to large numbers of low-capacity devices. Managing this flexibility is often driven by machine learning, and requires real-time control and aggregation of these devices, involving a diverse set of companies and devices and creating a longer chain of trust. This poses a security risk, as it is sensitive to adversarial machine learning, whereby models are fooled through malicious input, either for financial gain or to cause system disruption. We show the feasibility of such an attack by analysing empirical data of a real system, and propose directions for future research related to detection and defence mechanisms for these kind of attacks.

Energy Theft in Smart Grids: A Survey on Data-Driven Attack Strategies and Detection Methods

“…Nonetheless, a number of approaches have been proposed to confront complexity constraints through ramp strategies [17]. In parallel, the integration of DRES involves diverse types of data communication and systemon-chip technologies that are commonly manufactured with minimal security [23], [24]. Hence, enlarging the spectrum of cyber attacks that could be initiated such as to support potential energy theft acts [25].…”

“…Consumers or prosumers are also capable to lie on their demand data by utilising FDI techniques to cause under or over-reporting of energy consumption [23], [63]- [66]. We denote as γ i (t) to be the theft coefficient of node i at time t and O to be the set of consumers or prosumers providing falsified demand request data, where O = P ∩Q.…”

“…Other examples, include a combination of physical tampering of meters at various power grid levels (e.g., T&D, end-user smart-meters) where an attacker could identify through simple social engineering and bypassing of authentication protocols through ANSI optical ports with software such as Terminator that enables access [20]. In parallel, sophisticated MITM and FDI techniques may also consider the overall topology of a given grid deployment [72] in order to bypass any detection mechanisms whereas other utilise adversarial machine learning in order to game optimisation, scheduling and control processes within the EMS [17], [23]. The aforementioned technique is relatively new and exploits the deficiencies of automated management functions by manipulating and crafting falsified training data to machine learning-based processes that profile several measurements (e.g., ramp rate, power factor, reactive power) [23].…”

“…In parallel, sophisticated MITM and FDI techniques may also consider the overall topology of a given grid deployment [72] in order to bypass any detection mechanisms whereas other utilise adversarial machine learning in order to game optimisation, scheduling and control processes within the EMS [17], [23]. The aforementioned technique is relatively new and exploits the deficiencies of automated management functions by manipulating and crafting falsified training data to machine learning-based processes that profile several measurements (e.g., ramp rate, power factor, reactive power) [23].…”

“…For example, carrying out an attack where input data is made to look like normal electrical data, i.e., crafting an attack that seems normal to the machine learning algorithm or changing the weights of the trained ML model. These scenarios can maximize the predicted loss or falsify trained models to new adversarial objectives [23], [101], [102]. Moreover, handcrafted rulebased attacks are more sophisticated (than automated attacks)…”