There is without substitute in terms of security and confidence to run a complete node.
There are different concepts of "thin client". Some of them are…
BIP37 (flowering filter):
- [minus] With the current false positives rates used, peers can learn all the wallet addresses
- [minus] In general, it is carried out on an unencrypted channel (p2p 8333), ISP, etc. Also learn all your addresses
- [plus] the client can validate if the transactions are in a block (merkleblock)
- [plus] customers maintain a chain of blocks with headers can only at least verify PoW
- [plus] use only small bandwidth
- [minus] It is possible to hide transactions
- [plus]"Impossible" to falsify a transaction
Neutrino (compact block filters BIP158):
- [plus] Less privacy implications than BIP37 since filtering is done locally
- [minus] greater bandwidth consumption because the complete blocks must be downloaded (instead of Merkle "blocks" in BIP37)
- [minus] It is still possible to hide the transactions (although it is more complicated) because the block filters are not committed to the blocks (it would require a soft fork). The means are not compromised, the partners can falsify filters and cause them to lose the relevant transactions (can be [partially] mitigated by the comparison of filters of different pairs)
- [minus] There is no solution for mempool filtering (reliable "inbound transactions" can not be displayed)
- [plus] "Impossible" to falsify a transaction
Centralized validation (Bitpay, Samourai, etc.)
- [minus] Total trust in the company / server (they know all their addresses)
- [minus] You can hide overdue transactions
- [minus] You can create transactions artificially
- [plus] Minimum bandwidth consumption
If you want to see the chain of blocks without trusted third parties, you can must run a complete node (It could be removed with a space requirement of <10 GB, but Lightning implementations are not yet fully supported).
If and only if) BIP158 get block filters committed (Even if it's a soft fork, hash in base blocks of similar ones), it would no longer be possible to hide transactions through pairs, provide filters.