The most feared online entities in the present day are trolls. Trolls, a fearsome bunch of fake or pseudo online profiles, tend to attack online users, mostly celebrities, sports person or political profiles using a wide range of methods. One of these methods is to post obscene or NSFW (Not Safe For Work) content on your profile or website where User Generated Content (USG) is allowed. This can create unnecessary attention and cause legal troubles for you too.
The traditional way out is to get a moderator (or a team of them). Let all the USGs pass through this moderation system. This is a sustainable solution for a small platform. But if you are running a large scale app, say a publishing app where you publish one hundred stories a day, and the success of these stories depend on the user interaction with them, then this model of manual moderation becomes unsustainable. More the number of USGs, more is the turn-around time, larger the moderation team size. This results in escalating costs, for a purpose that’s not contributing to your business growth in any manner.
That’s where Machine Learning could help. Machine Learning algorithms that can scan images and content for possible abusive or adult content is a better solution that manual moderation. Tech giants like Microsoft, Google, Amazon have a ready solution for this. These companies have created APIs which are commercially available for developers. You can incorporate these APIs in your application to weed out the filth served by the trolls. The different APIs available for this purpose are Microsoft moderation, Google Vision, AWS Rekognition & Clarifai.
Dataturks have made a comparative study on using these APIs on one particular dataset to measure their efficiency. They used a YACVID dataset with 180 images, manually labelled 90 of these images as nude and the rest as non-nude. The dataset was then fed to the 4 APIs mentioned above, their efficiency was tested based on the following parameters.
- True Positive (TP): Given a safe photo, the API correctly says so
- False Positive (FP): Given an explicit photo but the API incorrectly classifies it as safe.
- False negative (FN): Given a safe photo but the API is not able to detect so and
- True negative(TN): Given an explicit photo and the API correctly says so.
TP and TN are two cases which meant the system behaved correctly. An FP meant that the app was vulnerable to attacks from trolls, FN meant the efficiency of the systems were low and hence not practically viable. 10% of the cases would be such that the API can’t decide whether its explicit or not. Those would be sent for manual moderation. This would bring down the maintenance cost of the moderation team.
The results that they received are shown below:
As it is evident from the above table, the best standalone API is Google vision with a 99% accuracy and 94% recall value. Recall value implies that if the same images are repeated, it can recognize them with 94% precision. The best results however were received with the combination of Microsoft and Google. The comparison of the response times are mentioned below:
The response time might have been affected with the fact that all the images accessed by the APIs were stored in Amazon S3. Hence AWS API might have had an unfair advantage on the response time. The timings were noted for 180 image calls per API.
The cost is the lowest for AWS Rekognition – $1 for 1000 calls to the API. It’s $1.2 for Clarifai, $1.5 for both Microsoft and Google. The one notable drawback of the Amazon API was that the images had to be stored as S3 objects, or converted into that. All the other APIs accepted any web links as possible source of images.
What this study says is that the power of filtering out negative and explicit content in your app is much easier now. You might still have to have a small team of moderators, but their jobs will be made a lot easier with the ML models implemented in these APIs.
Machine Learning is paving the way for us to be safe from the increasing menace of Trolls, a threat to free speech and open sharing of ideas which were the founding stones of internet and the world wide web as a whole. Will this discourage Trolls from continuing their slandering or will it create a counter system to bypass the APIs and checks? We can only know in time.