Illustrative examples highlighting impact of method on privacy
guarantees:
1. Laplace Noise vs. Gaussian Noise: Laplace noise addition is commonly
used for differential privacy. If two queries are executed sequentially,
each with Laplace noise addition, the cumulative privacy loss can be
calculated using sequential composition. In contrast, Gaussian noise
addition, while less common, could have a different impact on privacy
bounds due to its distinct noise distribution.
2. Exponential Mechanism: The exponential mechanism introduces
randomness to select outputs that maximize utility. If two queries use
the exponential mechanism in parallel, their combined effect is
determined by parallel composition. The choice of mechanism here
influences how parallel composition treats different queries.
3. Dynamic Sensitivity Mechanism: Differential privacy mechanisms like
the dynamic sensitivity mechanism adapt noise levels based on query
sensitivity. This dynamic behavior can interact with advanced
composition theorems, influencing how tighter bounds are achieved.
4. Adaptive Noise Mechanisms: Adaptive noise mechanisms adjust the
amount of noise based on the data and previous outcomes. The interplay
between adaptive noise mechanisms and advanced composition techniques
requires careful analysis to ensure the cumulative privacy guarantees
hold.
Overall, the relationship between composition techniques and
differential privacy mechanisms is intricate. The choice of mechanism
directly affects the application of composition techniques and can
impact the accuracy of privacy guarantees. Organizations must understand
these interactions and tailor their approaches to achieve the desired
level of privacy protection while optimizing utility.
Table 2.Comparative Study of Popular Composition
Algorithms12, 13