SparkCognition’s founder and CEO Amir Husain believes every tech company, especially in the AI field, should have a founding mission directing its strategy and evolution. In this case, the idea is that AI can only succeed in delivering value on broad scale beyond large enterprises with great resources if the models can build themselves. “At the core of our company is the idea that AI can build AI,” he extolls.
We have heard such mantras before, and certainly SparkCognition is let down by a website and promotional literature that overplay the AI and ML card, while in the same breath deploring such practice among its competitors. It warns against false rivals pretending they offer AI technology when they do not. This is a pity because when this marketing veneer is peeled away, we encounter a start-up that is making progress and to some extent living up to that founding philosophy.
This has been demonstrated by adapting the underlying technology to different sectors and problems, first targeting energy and industrial just after its launch in April 2014 in early stage mode. Then when it started attracting second stage finance, now standing at $56.5mn, the company diversified into defense and national security before entering finance as its third area.
This has taken it into predictive maintenance, cyber security monitoring and unstructured data analysis, served by four products. The first is Sparkpredict for predictive maintenance, where according to Husain the aim has been to “extend the window of forewarning” of failure so that operations such as oil pipelines or assembly plants have longer to take proactive actions like ordering a relevant replacement part.
The second product is Deeparmor with the focus on identifying malware or other agents of attack through their activity rather than being confined to specific signatures. It is easy for malware producers to evade signature detection just by changing a few incidental aspects of the software, rather like how the influenza virus can bypass human adaptive immunity by altering the shape of a few proteins on its coat.
The company also argues its ML approach training the system to identity patterns associated with malware activity is superior to more advanced heuristic techniques that rely on specific snippets of code for identification. In that case if the protection software fails to identify a particular action the malware will evade detection.
Heuristics also fails to cater for various obfuscation techniques on the part of the malware creator, such as being embedded within an encrypted file. An ML approach can even identify encrypted malware because it is just looking for unusual patterns and not specific items of code or data. The company cites the case of CryptoMix, a ransomware application that encrypts files on a victim’s machine and then demands payment to retrieve the information. The antivirus comparison website VirusTotal found that the number of antivirus vendors capable of detecting CryptoMix went down by 17% if just a single byte of its code was changed.
The third product, DeepNLP, is designed to analyze and extract information from unstructured data, which again can help with predictive maintenance as well as cybersecurity. Finally, Darwin is the component designed to help build ML models themselves with the help of expertise relating to a given target sector. This is the most unproven component whose success will determine whether the company really has a unique flavor.