Using neural nets for this purpose makes it likely to be susceptible to adversarial examples. Fairly harmless in this use case, but if a similar system would be

Using neural nets for this purpose makes it likely to be susceptible to adversarial examples. Fairly harmless in this use case, but if a similar system would be used for ContentID on YouTube, that could be exploited to subvert it (i.e. spam it with false positives and contest the verdicts en masse).