Communication between unmanned aerial vehicle (UAV) swarms and flying ad-hoc networks is critical and challenging since it is over open radio networks and energy, memory and processing budgets are severely constrained. Lightweight cryptography is the solution and ASCON — the NIST standard for lightweight authenticated encryption — has become a popular option for securing UAV links at a fraction of the energy cost of AES. The security of ASCON relies on one nonlinear device — 5-bit substitution box (S-box). However, this S-box is fixed and publicly known, and allows an adversary to precompute the cryptographic properties one time and then reuse them against all the protected traffic, and moreover, it does not meet the strict avalanche criterion (SAC) with an average SAC of 0.62, against the ideal value of 0.5. This paper proposes an S-box for ASCON with a key-dependent construction which can solve both problems. The secret key is used as an input for the deterministic generation of a byte stream, which is then input to a Fisher–Yates shuffle to generate a candidate for the random permutation of the thirty-two values, and a dual acceptance filter rejects any candidate that has a differential uniformity of more than 8, or a linearity of more than 16, which is the value of the ASCON's box. This process is continued until a candidate passes, thus each box emitted by the sender is guaranteed to be within these bounds, and the receiver will be able to perfectly reproduce the same box using the common key. This is executed on a set of 100 boxes, each key different. All these generated boxes are bijective and satisfy the differential, linear thresholds; the mean differential probability is improved from 0.25 to 0.225, and the mean dependence-matrix offset is improved from 0.2000 to 0.0932 as compared to the original ASCON box, while the S-box-level strict avalanche criterion is improved in each case. It is proved that the modified cipher possesses the ideal avalanche ratio of 0.5 from round 4 and onwards by full-permutation avalanche measurements, which is the same value as standard ASCON. The generation cost measured is a few milliseconds per session, with an average of 17 ms per session for a Raspberry Pi 5 class UAV companion computer, which is very feasible on constrained aero platforms.
