# Why the definition in $\epsilon$-differential privacy is multiplicative rather than additive？

According to its mathematical definition, a random algorithm $$M: D\rightarrow R$$ satisfies $$\epsilon$$-differential privacy if the adjacent datasets $$x, y \in D$$ where $$D$$ is a whole dataset and datasets $$x$$ and $$y$$ differs by only one record, and any set of $$S \in Range(M)$$, $$Pr(M(x) \in S) \leq Pr(M(y) \in S) * e^{\epsilon}$$.

The additive one is shown in this question: Differential privacy definition.

Dr. Dwork explains the advantage of using the multiplicative definition over the additive one in the Microsoft Research Lecture 2 (at about 2'50''). In short, with this multiplicative definition, it could be ruled out the possibility that an individual's record would be randomly selected and published.
However, I struggle to understand her meaning. I would appreciate any help in understanding this definition!

• DP implies additive guarantees, but the reverse is not true. In the textbook (shared in the answers below) 'The complexity of DP', this is discussed in detail. Commented Dec 12, 2020 at 22:56

In short, with this multiplicative definition, it could be ruled out the possibility that an individual's record would be randomly selected and published.

Consider a malicious algorithm $$M^*$$ that picks a random individual's record from the input database (of size $$n$$) and outputs it. Note that this $$M^*$$ should not be considered secure for a good differential privacy definition because it compromises the privacy of a random individual.

However, the additive differential privacy definition regards $$M^*$$ as a secure algorithm for $$\epsilon=1/n$$. To see this, consider two input databases $$X,Y$$ (both of size $$n$$) differ with a single element $$d$$ (where $$d\in X$$ but $$d\not\in Y$$). Then, it's not hard to see $$\Pr[M^*(X)\in S]\leq\Pr[M^*(Y)\in S]+1/n$$ holds for any subset $$S\subseteq R$$, i.e., $$M^*$$ satisfies the additive $$1/n$$-differential privacy.

The above issue is fixed by the standard multiplicative definition. Note that for $$S=\{d\}$$, there is no $$\epsilon$$ such that $$\Pr[M^*(X)\in S]\leq\Pr[M^*(Y)\in S]\cdot e^\epsilon$$ because $$\Pr[M^*(X)=d]=1/n>0$$ and $$\Pr[M^*(Y)=d]=0$$. That is, $$M^*$$ does not satisfy the standard $$\epsilon$$-differential privacy.

• I think it is exactly what Dwork mean in the video! Many thanks for your answer! Commented Mar 25, 2019 at 14:42

In Salil Vadhan's textbook The Complexity of Differential Privacy the author states in section 1.4

The choice of a multiplicative measure of closeness between distributions is important, and we will discuss the reasons for it later. It is technically more convenient to use $$e^\epsilon$$ instead of $$(1 + \epsilon)$$, because the former behaves more nicely under multiplication $$e^{\epsilon_1} \cdot e^{\epsilon_2} = e^{\epsilon_1 + \epsilon_2}$$.

You can find a very nice general discussion of why the definition is the way it is in Section 1.6.

One advantage of using $$e^\epsilon$$ instead of $$(1+\epsilon)$$ as far as I understand is that the composition guarantees (see Lemma 2.2) behave nicely. For instance if you have $$\epsilon_1$$-differentially private mechanism $$M_1$$ and $$\epsilon_2$$-differentially private mechanism $$M_2$$, then the mechanism $$M(x) = (M_1(x), M_2(x))$$ is $$(\epsilon_1 + \epsilon_2)$$-differentially private.

• Though your answer is different from the explanation Dwork gave, it explains from another practical aspect why the multiplicative definition is necessary. Thanks for your answer as well as the book! Commented Mar 25, 2019 at 14:41