Noa Garcia presents Unlearning in Image Generation: Can Concepts Be Erased?

While image generation offers great possibilities for data synthesis, it is
also surrounded by social concerns, often functioning as a machine that
reinforces existing stereotypes. Beyond bias, the field faces other challenges,
including issues related to data, intellectual property, and privacy. Although
technical solutions alone cannot solve problems embedded in society, they can at
least aim not to exacerbate them. Machine unlearning provides a technical
paradigm for addressing these and related concerns. However, are so-called
concept erasure methods based on machine unlearning truly ready? By examining
these methods beyond simple targeted prompts, we find that current approaches
remove concepts only at the surface level and the erased concepts remain
accessible when image generators are prompted with less explicit or more
indirect language.