Latest / EDGE AI POD / Got Fake Chips? Our AI Doesn't Fall For That

Transcript

Got Fake Chips? Our AI Doesn't Fall For That · EDGE AI POD

• 0:07
  Speaker 1: So, hi everyone, welcome to the talk.
• 0:08
  Speaker 1: Edge ML for counterfeit chip detection.
• 0:12
  Speaker 1: So we're both machine learning engineers in advanced R&D team in analog devices and we work on emerging security technologies and research method to utilize AI to enhance security and also work on solution to provide security for AI on the edge.
• 0:33
  Speaker 1: So here's today's agenda we're going to talk about permanent definition and we're going to give a brief summary of our solution, and then we're going to talk about details about our dataset collection and model architecture, and after that we're going to talk about results and roadmap.
• 0:51
  Speaker 1: So the semiconductor industry has grown into $500 billion market over the last 60 years.
• 1:01
  Speaker 1: However, there are some challenges, such as profound shortage of new chips and increase of counterfeit chips.
• 1:08
  Speaker 1: So examples of counterfeiting includes altering part markings to misrepresent the part, or presenting used part as new, or producing devices without authority, which leads to substantial risk of malfunction.
• 1:23
  Speaker 1: And, according to ERAI, the number of counterfeit chips is still increasing in the market, and counterfeit electronics cost global semiconductor industry more than $200 billion per year.
• 1:40
  Speaker 1: And so there are some existing works addressing this issue.
• 1:48
  Speaker 1: Although most of them successfully distinguish ICES, they have some limitations, such as they require hardware modification or complex setups or sometimes very sensitive to environmental noise.
• 2:03
  Speaker 1: Also, most of them utilize supervised machine learning algorithms, which requires the labeling of large corpus of data.
• 2:12
  Speaker 1: So here arises a need for an efficient counterfeit detection solution and we propose counterfeit detection solution that utilize unsupervised machine learning algorithm and analog fingerprint data.
• 2:29
  Speaker 1: And our method could be combined with traditional cryptographic authentication methods and we utilize very tiny machine learning models.
• 2:41
  Speaker 1: That requires less than 60 kilobytes of memory, so that is easy to incorporate into existing method.
• 2:47
  Speaker 1: Also, this method is very scalable and cost-effective.
• 2:51
  Speaker 1: So this is overview of deployment workflow.
• 2:57
  Speaker 1: So for model deployment we utilize the lightweight MCU with Cortex-M4 processor and we fuse pre-trained quantized model on board.
• 3:09
  Speaker 1: So here's how it works.
• 3:11
  Speaker 1: So from the chip we collect the real-time analog fingerprints, which are power of memory, write sequences and then we feed data to model and conduct inference on the board, and then we utilize this inference result to distinguish the authentic part or cloned part.
• 3:32
  Speaker 1: So from here I'm going to hand over to my teammate Pavani and she's going to give you further details.
• 3:42
  Speaker 2: Hey everyone, I'm going to give an overview of how we collected.
• 3:46
  Speaker 2: So this is an overview of how we collected the data set for our model.
• 3:59
  Speaker 2: So we chose to use our one-wire secure authenticator parts.
• 4:04
  Speaker 2: The reason behind using our one-wire secure authenticators for this solution is we know for a fact that some of these are getting cloned and it's a critical issue for our customers and we'll be solving that problem for them if we are able to differentiate between our authentic chip and somebody else's chip using our mm model.
• 4:23
  Speaker 2: So we took 1000 ds28e54 secure authenticators and offered 750 are in our training set and 250 are in the test set and our clone clone devices are closest relatives of DS28E54.
• 4:38
  Speaker 2: The reasoning behind choosing the closest relative of DS28E54 as clone is that if my model is able to tell me that this is a clone for a closest relative, it will do really good on some cheap microbe which has been done on other fab and everything right.
• 5:02
  Speaker 2: So we took all these samples and collected analog fingerprints while powering up this device and also during a memory write operation.
• 5:12
  Speaker 2: So we used logic analyzer to collect this analog fingerprints, digitizing it.
• 5:18
  Speaker 2: Later we downsample all this data that is collected to 1024 data points for each sample so that it's the less memory footprint for our ML model, so that we can use this on edge AI.
• 5:32
  Speaker 2: So I'll give a little bit overview of how the model architecture is and go through on like how the performance is for the current data set.
• 5:43
  Speaker 2: So the inspiration for this architecture is through auto encoder architecture for anomaly detection.
• 5:50
  Speaker 2: So one might think that how does auto encoder even differentiate between authentic and clones, right?
• 5:57
  Speaker 2: So I am training my autoencoder model using all our authentic parts and my model know really well to reconstruct the similar signature for my parts.
• 6:08
  Speaker 2: But in case, if it comes with the cloner part, it doesn't know how to do it so it has a larger reconstruction loss.
• 6:15
  Speaker 2: So that is my identifying parameter to differentiate between authentic and clones.
• 6:21
  Speaker 2: So we went through multiple iterations on model architecture, depth and also what kind of layers that we want to use.
• 6:28
  Speaker 2: We start off with single channel, like sending power of samples first across some convolution layers, but that was not really working out.
• 6:37
  Speaker 2: And then we used two-channel convolutional autoencoder model.
• 6:41
  Speaker 2: That was giving a really good performance.
• 6:43
  Speaker 2: But when I add my application code to collect the real-time data it is too big for the micro that we were using.
• 6:50
  Speaker 2: So later we used convolution 2D layers, which reduced the model size better and we were able to fuse it into the micro.
• 7:00
  Speaker 2: So we trained the model for like 100 epochs and observed really good convergence and also really good losses.
• 7:09
  Speaker 2: And, based on the threshold that comes from the pre-trained model, we tested our data set that we curated before, with some of the clones and some of the authentic, and the model was able to 100% differentiate on the existing data set.
• 7:26
  Speaker 2: And this this is basically another performance metric for how well the distribution is separated.
• 7:34
  Speaker 2: So so the distance ratio for the current model is 0.75.
• 7:39
  Speaker 2: So it clearly shows the difference between our authentic samples and the clone samples, which gives us more confidence while differentiating them.
• 7:49
  Speaker 2: So, to summarize, we have collected the data, built the model and deployed it on Cortex-M4.
• 7:56
  Speaker 2: Further, we were planning to deploy this model onto a processor which has a CNN accelerator and see how the power and latency changes happen, and also introduce this model in different environment conditions and process variation and make it more robust.
• 8:15
  Speaker 2: And one more thing that we also want to try is automate the data collection so that one can collect these analog fingerprints directly from the AT equipment that they have.
• 8:28
  Speaker 2: So I want to leave everybody with this thought.
• 8:33
  Speaker 2: This solution is not just limited to secure authenticator parts.
• 8:38
  Speaker 2: Right, if you are able to collect any analog fingerprints, you are still able to train your unsupervised model and implement it, so it is more scalable and reusable too.
• 8:50
  Speaker 2: The solution would always be like an add-on to existing cryptographic authentications as well, because just the talk before they have mentioned a lot of side channel attacks and your keys can be confiscated.
• 9:04
  Speaker 2: So in scenarios like that, if you have an another technique to identify your own chip, that is a win-win.
• 9:11
  Speaker 2: And this solution is also a softwarebased too and can be implemented into different kinds of micros.
• 9:20
  Speaker 2: As of now, we are using Cortex-M4, but we can still make changes and put it on existing micros as well.
• 9:27
  Speaker 2: So with this, I'm going to open it for Q&A.
• 9:29
  Speaker 2: Thank you so much.
• 9:32
  Speaker 1: Let's thank the speakers Selena and Bhavana.
• 9:34
  Speaker 1: Thank you very much.