Machine learning for automatic identification of points of interest for side channel leakage
Methods, machine readable media and systems for evaluating, through one or more simulations, the leakage of sensitive data in an integrated circuit, such as cryptographic data or keys, are described. The embodiments can use machine learning models, such as one or more neural networks to generate one or more leakage related scores for each portion in a set of portions of the cryptographic data. In one embodiment, leakage data associated the first set of POIs with one or more neural networks is processed by the one or more neural networks to identify the POIs that leak the most and determine one or more scores for each portion in the set of portions of the cryptographic data.
1 . A non-transitory machine readable medium storing executable program instructions which when executed by a data processing system cause the data processing system to perform a method, the method comprising:
simulating a circuit processing cryptographic data to obtain a set of points of interest (POIs) in the simulated circuit, the cryptographic data including a plurality portions, and the set of POIs being associated with leakage data from the processing of the cryptographic data;
determining, within each POI in the set of POIs, a coupling impact of neighboring tiles relative to a center tile in each POI;
selecting a fusion pattern for each POI based on the determined coupling impact, the fusion pattern for each POI is used to statistically aggregate leakage data from each tile in the fusion pattern; and
processing the leakage data with a plurality of neural networks, each neural network corresponding to a respective portion of the cryptographic data, the plurality of neural networks receiving the leakage data as inputs to predict probabilities of leakage for corresponding portions of the cryptographic data.
2 . The non-transitory machine readable medium as in claim 1 , wherein the method further comprises:
filtering the leakage data from the processing of the cryptographic data to derive the set of POIs which is less than all POIs found in the circuit.
3 . The non-transitory machine readable medium as in claim 2 , wherein the filtering comprises an image processing edge detection.
4 . The non-transitory machine readable medium as in claim 1 , wherein the simulating the circuit comprises:
providing through direct vector control at least a portion of the cryptographic data.
5 . The non-transitory machine readable medium as in claim 1 , wherein the leakage data comprises at least one of:
power noise data;
thermal data;
electromagnetic emission data; or
tile-based power data when each POI in the set of POIs comprises a plurality of tiles,
and wherein the leakage data includes temporal samplings during processing of the cryptographic data and spatial samplings over the circuit.
6 . The non-transitory machine readable medium as in claim 1 , wherein the plurality of neural networks comprise a first set of N neural networks, wherein the cryptographic data comprises N portions, and wherein each of the N neural networks is trained to identify a highest leakage POI for a corresponding portion of the cryptographic data.
7 . The non-transitory machine readable medium as in claim 6 , wherein the plurality of neural networks further comprise a second set of neural networks that receive leakage data of a plurality of the highest leakage POIs identified by the first set of N neural networks, and wherein the second set of neural networks is trained to generate a sensitivity score indicating a leakage sensitivity of a corresponding POI for each of the N portions of the cryptographic data.
8 . The non-transitory machine readable medium as in claim 7 , wherein the second set of neural networks is further trained to generate a set of probability scores indicating a probability that each of the N portions of the cryptographic data will be leaked by the circuit.
9 . The non-transitory machine readable medium as in claim 7 , wherein the second set of neural networks is further trained to generate a ranked list of a plurality of the sensitivity scores for the set of POIs.
10 . A computer-implemented method for simulating leakage of sensitive data from a circuit, comprising:
simulating the circuit to process cryptographic data to obtain a set of points of interest (POIs) in the circuit, the cryptographic data including a plurality portions, and the set of POIs being associated with leakage data from the processing of the cryptographic data;
determining, within each POI in the set of POIs, a coupling impact of neighboring tiles relative to a center tile in each POI;
selecting a fusion pattern for each POI based on the determined coupling impact, the fusion pattern for each POI is used to statistically aggregate leakage data from each tile in the fusion pattern; and
processing the leakage data with a plurality of neural networks, each neural network corresponding to a respective portion of the cryptographic data, the plurality of neural networks receiving the leakage data as inputs to predict probabilities of leakage for corresponding portions of the cryptographic data.
11 . The method as in claim 10 , further comprising:
filtering the leakage data from the processing of the cryptographic data to derive the set of POIs which is less than all POIs found in the circuit.
12 . The method as in claim 11 , wherein the filtering comprises an image processing edge detection.
13 . The method as in claim 10 , wherein the simulating the circuit comprises:
providing through direct vector control at least a portion of the cryptographic data.
14 . The method as in claim 10 , wherein the leakage data comprises at least one of:
power noise data;
thermal data;
electromagnetic emission data; or
tile-based power data when each POI in the set of POIs comprises a plurality of tiles,
and wherein the leakage data includes temporal samplings during processing of the cryptographic data and spatial samplings over the circuit.
15 . The method as in claim 10 , wherein the plurality of neural networks comprise a first set of N neural networks, wherein the cryptographic data comprises N portions, and wherein each of the N neural networks is trained to identify a highest leakage POI for a corresponding portion of the cryptographic data.
16 . The method as in claim 15 , wherein the plurality of neural networks further comprise a second set of neural networks that receive leakage data of a plurality of the highest leakage POIs identified by the first set of N neural networks, and wherein the second set of neural networks is trained to generate a sensitivity score indicating a leakage sensitivity of a corresponding POI for each of the N portions of the cryptographic data.
17 . The method as in claim 16 , wherein the second set of neural networks is further trained to generate a set of probability scores indicating a probability that each of the N portions of the cryptographic data will be leaked by the circuit.
18 . The method as in claim 16 , wherein the second set of neural networks is further trained to generate a ranked list of a plurality of the sensitivity scores for the set of POIs.