Here is some food for thought – did you know that the formula for Coca Cola is not patented? Neither is Kentucky Fried Chicken’s original recipe. Neither are the algorithms used for the most significant patterns in web searches. The reason? In order to acquire a patent, one must explain a process and disclose it for all to see and challenge. Now, in the world of IT security, we secure the personal data of our customers. In some manufacturing processes, we secure the precision of temperatures for the proper melting points, etc. But like these examples, we need to think about the intellectual property (IP) that comes into play with the introduction of robotics and cognitive capabilities organizational processes. Not just the data, but the precision in which the decisions are made must also be secured.
In the world of AI and robotics, we need to think about how we secure the logic and more importantly the trained models of AI. This is because IP is not the model we run, but the results of the data running through that model. Ultimately, these models need to be secured to drive competitive advantage. Herein lies the issue. We can’t simply ship out 100 lbs. of mix and coat chicken with it; we need to look at all aspects of the IP.
There are four key things to be secured. The first is the most obvious: Data. The data of 100,000 instances where we have a result is extremely valuable. Specifically, if we have 100 columns and a single additional to tell us a particular situation, then the data holds the patterns. Interestingly enough, this is common to data scientists. What they may not tell you is that if you take out a random set of 10000 of those samples, the end result could change. The robustness of the values determine just how much. Secrets.
The second thing we need to protect is the transformation of the 100 columns of data into predictable values. How are the values normalized? Do we use certain combinations and execute “feature engineering”, or perhaps we consider conditional predictors and assist the learning? Secrets.
The third thing is the model itself. Specifically we could run any number of models (K-means, CNNs, Naïve Bayes, LSTMs, etc.). The sequence we run them in, the inputs they take, and the use of the outputs all create an artistry to the AI modeling. These sequences are equivalent to knowing whether to turn left or right and how many times on a combination lock. Secrets.
The final thing is how do these predictors going through the models provide the actionable results, and more significantly, allow us as humans to see what is the learning? We consider those dimensions of the outcome. What dimensions do we have and within those dimensions what attributes do we examine to see the outcomes and the influences of those outcomes. Secrets.
At Virtusa, we conceive security along the lines explained in this blog. We believe our customers too need insights on how to protect more than just the data. After all, the data is now creating the programs.
Your source code for your legacy systems aren’t open source, are they?