Reducing malicious use of synthetic media research: Considerations and potential release practices for machine learning