Let $H:X \to \{0,1\}^b$ denote a cryptographically secure, $b$-bits hash function on a set $X$. Let $H^∗:\mathcal P(X) \to \{0,1\}^b$ be a function on the power set of $X$ defined by $H^∗(\{x_1,…,x_n\}) = \Sigma_i H(x_i)$

Is see and understand that most variants of this construction using a modular domain for addition are flawed. I also see that these failures do not contradict the NP-completeness of the subset sum problem.

I am however not finding any arguments against the security of this scheme in the non modular case, where the sum is in the natural integer domain. At first glance, one would assume that any solution of the non modular variant that improves on the $\mathcal O(2^{n/2})$ bound can be turned into a an improved algorithm for the subset sum problem for random inputs.

Formally, the sum of hashes in the natural domain does not fit the definition of a hash function since the size of the output is variable. This is however easy to fix in practice by applying the hash function once more to the sum.

This is obviously related to previous questions on this forum, such as Is the sum of hashes a suitable hash for sets? and Are there any practical implementation of a homomorphic hashing or signature scheme?

  • 3
    $\begingroup$ If you apply the hash function once more, wouldn't that lose the homomorphic property? $\endgroup$
    – Maarten Bodewes
    Commented Jun 9 at 2:39
  • 1
    $\begingroup$ Touché! It does indeed :-) Still makes for an updatable hash. Another way around is to pad the hash sum to $3b/2$ bits in length and accept that it security parameter is still $k = b/2$. $\endgroup$
    – Philippe
    Commented Jun 9 at 11:57
  • $\begingroup$ This simple comment by @maarten-bodewes is making me realize that the security parameter of $H^*$ is not as directly related to the security parameter of $H$ as I assumed it should be. If the input set to $H^*$ is expected to be up to $2^l$ in cardinality, to achieve a $H^*$ with security parameter $k$, you need $H$ to start with a security parameter of $l+k$. If you are willing to perform $2^l$ invocations of $H$, you must force your adversary to perform the equivalent of $2^{l+k}$ invocations in his attack. $\endgroup$
    – Philippe
    Commented Jun 9 at 13:36
  • $\begingroup$ $H^*$ is not collision-free. There are simple collisions like $$H^∗(\{x_1,…,x_n\}) = H^∗(\{x_n,…,x_1\}) = \Sigma_i H(x_i)$$ $\endgroup$
    – kelalaka
    Commented Jun 9 at 21:19
  • 1
    $\begingroup$ The problem of finding a collision is closely related to the Signed Subset Sum Problem, a variation of the Knapsack problem. $\endgroup$
    – fgrieu
    Commented Jun 12 at 4:54

2 Answers 2


The issue here is that even if the distribution of $H$ is uniform (which is the ideal that we aim for), the distribution of $\sum H$ is not. Thus if $n=2$ the distribution is triangular and as $n$ grows the distribution tends towards $\mathcal N(n2^{b-1},O(n4^{b-1}))$ per the central limit theorem. From a security point of view then the collision entropy decreases hugely in comparison to the max-entropy (Hartley entropy).

Consider for example the case $n=1200$, which we can approximate as normally distributed with mean $1200\cdot 2^{b-1}$ and standard deviation $10\cdot 2^{b-1}$. Even naively, we may generate random collections of 1200 sets and evaluate our sum of hashes function of these random collections. We'll discard sets where the sum of hashes lies outside $[1140\cdot 2^{b-1},1260\cdot 2^{b-1}]$ which is a $\approx 2\sigma$ interval and allows us to keep $\approx 95\%$ of our sets. Our interval contains $80\cdot 2^{b-1}$ hash values, and even with a naive bound we would expect a collision from $n=\sqrt{40\cdot 2^{b-1}}/0.95\approx 6.65\times 2^{(b-1)/2}$ trials which is below the desired bound of $\sqrt{\pi\cdot 1200\cdot 2^{b-1}}\approx 108.8\times2^{(b-1)/2}$.

  • $\begingroup$ And in search of colliding sets, we need not choose the elements in the sets naively. In particular, when we hash a new message, we can add it to a set selected such that the updated hash is close to $n2^{b-1}$, thus tweaking the distribution of the hashes we keep to increase the probability of collision. $\endgroup$
    – fgrieu
    Commented Jun 11 at 3:52
  • 1
    $\begingroup$ @Philippe It should indeed; I've now corrected. $\endgroup$
    – Daniel S
    Commented Jun 11 at 6:46
  • $\begingroup$ I am not sure the argument works. With the standard deviation still dominated by the $2^b$ factor, it looks like any gain is related to the amount of work done (in filtering samples or skewing the distribution) in ways that are already factored into the choice of $b$. The advantage shown in @Daniels S' calculation is near the expected $\sqrt{1200} = 34.641$. I think I may need to make the definition of 'security' more explicit since the security of $H^*$ is necessarily weaker than that of $H$. $\endgroup$
    – Philippe
    Commented Jun 11 at 20:51

I believe this shows that the scheme is a lot less secure than the original hash function $H$, with a security parameter only proportional to the square root of the output size of $H$, in the ideal case. It is a relatively straight forward adaptation of Wagner2002. It shows that it is not, as I was hoping, the use of modular groups that make the simple additive methods ineffective. Non modular addition is just as vulnerable, it seems.

Let $\mathcal{Y}$ be the set of triplets of the form $(\mathcal{h},\mathcal{A},\mathcal{B})$ where:

  • $\mathcal{A}$, $\mathcal{B}$ are subsets of $X$
  • $H^*(\mathcal{A}) \ge H^*(\mathcal{B})$
  • $\mathcal{h} = H^*(\mathcal{A}) - H^*(\mathcal{B})$, a non negative integer

We define a combining binary operator $\oplus \colon \mathcal{Y} \times \mathcal{Y} \to \mathcal{Y}$. Given two elements of $\mathcal{Y}$, say $y_r = (\mathcal{h}_r,\mathcal{A}_r,\mathcal{B}_r)$ and $y_s = (\mathcal{h}_s,\mathcal{A}_s,\mathcal{B}_s)$, with $\mathcal{A}_r$, $\mathcal{B}_r$, $\mathcal{A}_s$ and $\mathcal{B}_s$ all disjoint:

  • if $\mathcal{h}_r < \mathcal{h}_s$, then $y_r \oplus y_s \to y_s \oplus y_r$
  • otherwise $y_r \oplus y_s \to (\mathcal{h}_r - \mathcal{h}_s, \mathcal{A}_r \cup \mathcal{B}_s, \mathcal{A}_s \cup \mathcal{B}_r)$

The idea is that the elements of $\mathcal{Y}$ are collision candidates for $H^*$, and the $\oplus$ operator combines them to form a new candidate.

We create a stream $S^0$ of elements from $\mathcal{Y}$ by enumerating elements of $X$ in some order, and setting the $i^{th}$ element of $S^0$ to $ S^0_i = (H(x_i), \{x_i\}, \emptyset)$.

We build a new stream $S^1$ by consuming elements of $S^0$ and storing them in a hash table indexed by the $l$ lower order bits of the hash difference. When a collision is found, say $S^0_i = (\mathcal{h}^0_i,\mathcal{A}^0_i,\mathcal{B}^0_i)$ and $S^0_j = (\mathcal{h}^0_j,\mathcal{A}^0_j,\mathcal{B}^0_j)$, where $\mathcal{h}^0_i$ coincides with $\mathcal{h}^0_j$ in their $l$ low-order bits, we output $S^0_i \oplus S^0_j$ and erase $S^0_i$ and $S^0_j$ from the hash table.

We build new streams, $S^k$, by consuming elements of $S^{k-1}$ and storing them in a new hash table indexed by the $k\cdot l$ lower order bits of the hash difference. When a collision is found, say $S^{k-1}_i$ coincides with $S^{k-1}_j$ in the $k\cdot l$ lower order bits, we output $S^{k-1}_i \oplus S^{k-1}_j$ and erase $S^{k-1}_i$ and $S^{k-1}_j$ from the hash table.

Note that if $k > 0$, $S^k$ is a stream of collision candidates, made of triples $(\mathcal{h},\mathcal{A},\mathcal{B})$ where $\mathcal{A}$ and $\mathcal{B}$ each have cardinality $2^{k-1}$.

If $k\cdot l \ge b + k$, $S^k$ is a stream of actual collisions.

To get the first real collision, you need to fill in all $k$ hash tables. This will consume on average about $2^{k+l}$ elements of $S^0$. We therefore chose $l = \lceil \sqrt{b} \rceil$ and $k = \lceil (b + k) / l \rceil$. The running time is $\mathcal O(2^{2\sqrt{b}})$ invocations of $H$.

Thereafter, to get the next collision, you need to consume an average of $2^k$ elements of $S^0$. The running time is only $\mathcal O(2^\sqrt{b})$, just enough time to calculate $H^*$ on the next colliding pair.

The storage requirement is $\mathcal O(\sqrt{b} 2^\sqrt{b})$.

[ I expect this is not very intelligible. Am willing to update and clarify. ]


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.