Apparatus and method to detect power supply security attack and risk mitigation
Hardware based unsupervised based machine-learning (ML) approach to identify a security threat to the processor (e.g., caused by probing of a power supply rail). An apparatus is provided which includes an on-die power supply droop detector as a feature extractor. The droop detector detects a droop in the power supply caused by a probe physically coupling to the power supply rail. The droop detector in combination with machine-learning logic detects change in power supply rail impedance profile due to a probe coupled to the power supply rail. A deep-neural network (DNN) is provided for feature classification that classifies a security threat from normal operation and from operations caused by aging of devices in the processor. The DNN is trained in a training phase or production phase of the processor. An aging sensor is used to distinguish classification of aged data vs. normal data and data from security attack.
1. An apparatus, comprising:
a droop detector to sample voltages on a power supply rail during a droop on the power supply rail, wherein the voltages are sampled at different time points in a time window to determine a droop profile; and
a machine-learning logic coupled to the droop detector, wherein the machine-learning logic is to extract features of the droop profile and, based on the extracted features, determine whether to classify the droop as a security threat to a processor.
2. The apparatus of claim 1 , wherein the droop detector comprises a comparator, and the machine-learning logic is to adjust a gain of the comparator during unsupervised learning of the machine-learning logic.
3. The apparatus of claim 1 , further comprising a memory to store weights applied by the machine-learning logic to classify the droop.
4. The apparatus of claim 1 , wherein the machine-learning logic is part of an in-memory logic that comprises memory and computational blocks to determine weights and use the weights to classify the droop.
5. The apparatus of claim 1 , wherein the droop detector is to sample voltages on the power supply rail during different droops on the power supply rail, and the voltages of the different droops are sampled at different time points in different time windows to generate an array of data for the machine-learning logic to process.
6. The apparatus of claim 1 , further comprising a shift-register to store an output of the droop detector.
7. The apparatus of claim 1 , wherein: the droop detector comprises a comparator to compare the voltages on the power supply rail to an adjustable reference voltage, the apparatus further comprises a reference generator to provide the adjustable reference voltage to the comparator, and the machine-learning logic is to adjust the adjustable reference voltage.
8. The apparatus of claim 1 , wherein the machine-learning logic comprises a deep neural network to apply unsupervised machine-learning.
9. The apparatus of claim 1 , wherein:
the machine-learning logic is to separate an output of a neural network into clusters; and
to determine whether to classify the droop as the security threat, the machine-learning logic is to determine whether an inter-cluster distance of the neural network is greater than a threshold.
10. The apparatus of claim 9 , wherein:
the power rail is to power the processor; and
to determine whether to classify the droop as involving aging of devices in the processor, the machine-learning logic is to determine whether an intra-cluster distance of the neural network is greater than a threshold.
11. The apparatus of claim 1 , wherein the apparatus is in a system, and if the droop is classified as the security threat, the machine-learning logic is to power down the system.
12. The apparatus of claim 1 , further comprising an aging sensor coupled to the machine-learning logic, wherein the machine-learning logic is to determine whether to classify the droop as an aging-related event based on data from the aging sensor.
13. The apparatus of claim 1 , wherein to determine whether to classify the droop as the security threat:
the droop detector is to sample voltages on the power rail during different droops on the power supply rail, wherein the voltages of the different voltage droops are sampled at different time points in different time windows to determine different droop profiles; and
the machine-learning logic is to extract features of the different droop profiles.
14. The apparatus of claim 1 , wherein:
if the machine-learning logic does not classify the droop as the security threat, the machine-learning logic is to determine whether to classify the droop as an aging event; and
if the machine-learning logic does not classify the droop as the aging event, the machine-learning logic is to determine whether to classify the droop as a non-security event.
15. The apparatus of claim 1 , wherein:
the power rail is to power the processor; and
if the droop is classified as the security threat, the machine-learning logic is to disconnect the processor from a peripheral.
16. The apparatus of claim 1 , wherein if the machine-learning logic classifies the droop as the security threat, the machine-learning logic is to reduce access to a register.
17. An apparatus, comprising:
a voltage regulator coupled to a power supply rail;
an aging sensor to detect aging of one or more devices in the apparatus;
a droop detector to detect a droop of a voltage on the power supply rail; and
a machine-learning logic coupled to the droop detector and the aging sensor, wherein the machine-learning logic is to classify the droop as a security threat to the apparatus when the voltage has a first variation, a non-security event when the voltage has a second variation less than the first variation, or an aging event when the voltage has a variation which is within a range of the second variation.
18. The apparatus of claim 17 , wherein the droop detector is to sample the voltage at different times and for different time windows to generate an array of data for the machine-learning logic, and the machine-learning logic to classify the droop based on the array of data.
19. The apparatus of claim 17 , further comprising a memory to store weights applied by the machine-learning logic to classify the droop.
20. A system, comprising:
a memory; and
a processor coupled to the memory, wherein the processor comprises:
a droop detector to detect a droop profile of a voltage droop on a power supply rail; and
a machine-learning logic coupled to the droop detector, wherein the machine-learning logic is to:
extract features of the droop profile;
implement a neural network;
based on the extracted features, determine an inter-cluster distance of the neural network; and
determine whether to classify the voltage droop as a result of probing of the power supply rail based on the inter-cluster distance.
21. The system of claim 20 , wherein:
the power rail is to power the processor; and
to determine whether to classify the voltage droop as involving aging of devices in the processor, the machine-learning logic is to determine whether an intra-cluster distance of the neural network is greater than a threshold.