Acoustic pornography recognition using CNN and bag of refinements

I'm unfamiliar with 'pornography recognition' as an established task in ML research (lol), but for what it's worth, it's not an innovate use of CNNs for audio classification. You can essentially turn any audio classification task into an image problem (raw audio into features like spectrograms/MFCCs). Which people have been doing since forever (by which I mean a number of years now).

Porn filtering is useless or harmful?

It's like telling me you don't have kids without telling me you don't have kids xD.

I have my 8 year old sons internet pretty much locked down at this point, but people still end up skirting around the rules of sites like youtube when publishing content towards kids. Automatic filtering is one of the only ways to prevent newly published content geared towards children from being malicious. The alternative, human curation, even if done by method of reporting a video, is time consuming and expensive.

Whether you like porn or hate porn, classifying porn is useful. There is a lot of it around and it causes a lot of practical problems if you are hosting public content if you're not in control of it.

Porn has been proven to have severe detrimental physical and mental effects, including: erectile dysfunction, depression, changing perceptions of sex, etc.

Feel free to google them (going on google scholar and typing in “pornography negative impacts” is a good start), or I can cite a laundry list of studies saying the same thing.