Beyond flatness 2

Ever since social media was invented, it became a space for exercising freedom of speech. In the past few years, it has arisen even more powerfully as an effective tool for raising awareness and calling for social justice. Though it was created for the social good, it is never a pure land, especially when Internet censorship has recently become more stringent. As human-led and AI-driven keyword screening exerted by the governmental inspectorate is almost inevitable on social media, depriving room for diversity in public opinions, I thought of the possibility of utilizing Adversarial Attacks to stage resistance to the ubiquitous media censorship.

Beyond Flatness is an online platform helping social media users generate corresponding renderings of their inputted text. By downloading manipulated images from the platform, they can freely send these messages over the Internet without worrying about being banned and censored regularly. They can regain the information without any extra methods by tilting their personal devices to a specific angle.

I have run the outcome with text recognition software and done several tests on social media platforms. The result shows that machines have failed to detect the text embedded in those images. Between the provided deformed styles on the platform, an image similarity recognition software gives a score of 68 (out of 100), indicating even if an image contains the same text information, a style change can decrease the similarity to around 30%. Last but not least, if sending those images straight to an image recognition software, they are recognized as glitch arts or stripe patterns rather than distorted text images.

To go around the human-led and AI-driven screenings, I developed a generator with p5.js that produces artwork of distorted letters in different colors and shapes each time. Utilizing Adversarial theory to back up the tool allows the machine to perform Adversarial attacks on the algorithms and therefore defends the subjects from the same type of attacks. In the case of producing images and letters, it allows the generated artworks to bypass machine scanning as the information is unreadable to the computers. At the same time, if they contain sensitive keywords, these images require a lot of effort to review manually and thus dramatically slow down the process before they are banned from the Internet. This way, information carried through these images lasts longer on social media, allowing more time for creating iterations to survive.

To further understand how the Adversarial Attacks work in the process of image production, I experimented with two methods to interfere with AI Optical Character Recognition (OCR): Image Noise and Image Morph.

P5JS