Regarding VDFs, I want to bring up a term that I find important, “asymmetry”.
VDFs are good in terms of asymmetry, because the attacker pays 50x workload and you verify this just x. You benefit basically 49x against the attacker. In literature, VDFs are considered for use in consensus mechanisms with big inputs such as 1 second, which is meaningful for the benefits of this asymmetry. But in low delay operations, such as 50-100 ms around (if this is more for each hop, total libp2p-mix delay gets increased too much), this 49x benefits VDF’s symmetry is getting less beneficial since verifying VDF takes more time in terms of proportion than calculating. The spam attacker can abuse this advantage to send invalid VDF calculations to execute DOS attacks on the mixnet instances. This can be feasible since VDF calculations work in a single core, so with the rest of the free cores, the attacker can create more invalid VDFs.
As Prem stated here, RandomX cannot be suitable for mobile devices; Equihash can provide a fairer and easier solution among mobile and powerful devices because its similar high-end RAM of laptops and mobiles, also asymmetry here is far more than VDFs, so the attack above is not valid here. It’s ref implementation here.