Crafting canaries to measure empirical privacy of DP-FL training under a realistic threat model. - View it on GitHub